Highlights

A Survey on Neural-symbolic Learning Systems

Dongran Yu,Bo Yang,Dayou Liu,Hui Wang,Shirui Pan

• Neural-symbolic learning systems combine the neural systems and the symbolic systems into a unified
framework.

• Neural-symbolic learning systems can equip AI with the ability to perform perception and cognition.
arXiv:2111.08164v3 [cs.LG] 25 Jun 2023

• A good combination between the neural systems and the symbolic systems allows the model to achieve the
desired performance.
A Survey on Neural-symbolic Learning Systems
Dongran Yua,c , Bo Yanga,b,∗ , Dayou Liua,b , Hui Wangd and Shirui Pane
a School of Computer Science and Technology and the Key Laboratory of Symbolic Computation and Knowledge Engineer (Jilin University),
Ministry of Education, Changchun, Jilin 130012, China
b School of Computer Science and Technology, Jilin University, Changchun, Jilin, 130012, China
c School of Artificial Intelligence, Jilin University, Changchun, Jilin, 130012, China
d School of Electronics, Electrical Engineering and Computer Science, Queen’s University Belfast
e School of Information and Communication Technology, Griffith University

ARTICLE INFO ABSTRACT

Keywords: In recent years, neural systems have demonstrated highly effective learning ability and superior
Neural-symbolic Learning Systems perception intelligence. However, they have been found to lack effective reasoning and cognitive
Neural Networks ability. On the other hand, symbolic systems exhibit exceptional cognitive intelligence but suffer
Symbolic Reasoning from poor learning capabilities when compared to neural systems. Recognizing the advantages
Symbols and disadvantages of both methodologies, an ideal solution emerges: combining neural sys-
Logic tems and symbolic systems to create neural-symbolic learning systems that possess powerful
Knowledge Graphs perception and cognition. The purpose of this paper is to survey the advancements in neural-
symbolic learning systems from four distinct perspectives: challenges, methods, applications,
and future directions. By doing so, this research aims to propel this emerging field forward,
offering researchers a comprehensive and holistic overview. This overview will not only highlight
the current state-of-the-art but also identify promising avenues for future research.

1. Introduction
Perception, represented by connectionism or neural systems, and cognition, represented by symbolism or symbolic
systems, are two fundamental paradigms in the field of artificial intelligence (AI), each having prevailed for several
decades. Figure 1 showcases the rise and fall of these two doctrines, which researchers commonly categorize into
three significant periods. (1) The 1960s-1970s (1956-1968). The inception of reasoning in AI can be traced back to
the 1956 Dartmouth Conference, where Newell and Simon introduced the "logical theorist" program, successfully
proving 38 mathematical theorems. This marked the beginning of the reasoning era in AI. However, researchers soon
realized that relying solely on heuristic search algorithms had limitations, and many complex problems necessitated
specialized domain knowledge to achieve higher levels of intelligence. Consequently, incorporating knowledge into
AI models became a prevalent notion, as it enhanced the ability to find solutions within large solution spaces. (2) The
1970s-1990s (1968-1985). During this period, significant developments occurred, such as the creation of the first expert
system, "DenDral" by Faigan’s nanny and Lederberg in 1968. This milestone represented the organic integration of AI
and domain knowledge, marking the advent of the knowledge-focused era in AI. However, knowledge acquisition posed
a significant challenge, leading to a shift towards automatic acquisition of valuable knowledge from massive datasets,
which gradually became the mainstream trend in AI. (3) From the 1990s to the present: In 1983, neural networks
started gaining prominence, and after 2000, AI entered the era of machine learning. Notable breakthroughs in machine
learning, particularly through neural networks, include the 2012 triumph of ImageNet with Deep Convolutional Neural
Networks (CNN), and the 2016 victory of AlphaGo against the Go world champion. These significant milestones
exemplify the power of machine learning approaches, predominantly driven by neural networks.
To date, neural networks have demonstrated remarkable accomplishments in perception-related tasks, such as image
recognition [108]. However, there exist various scenarios, including question answering [48], medical diagnosis [111],
and autonomous driving [138], where relying solely on perception can present limitations or yield unsatisfactory
outcomes. For instance, when confronted with unseen situations during training, machines may struggle to make
accurate decisions in medical diagnosis. Another crucial consideration is the compatibility of purely perception-
based models with the principles of explainable AI [105]. Neural networks, being black-box systems, are unable to
∗ Correspondingauthor. School of Computer Science and Technology, Jilin University, Changchun, Jilin, 130012, China (Note:Accepted by
Neural Networks Journal)
[email protected] (D. Yu); [email protected] (B. Yang); [email protected] (S. Pan)

Dongran Yu et al.: Preprint submitted to Elsevier Page 1 of 29

 A Survey on Neural-symbolic Learning Systems

AlphaGo defeats the world Go champion

CNN handles ImageNet to win 2016

Learning
DENDRAL The world's first expert system was born 2012
Learning
The Dartmouth Conference gave 1968
birth to artificial intelligence science Expert
system
1956
Logic
Machine learning stage
Knowledge stage Computers learn algorithms from
Reasoning stage data, and deep learning is a great
Summarize human knowledge
Give logical reasoning ability to success in the field of speech and
and teach computer systems
computer system images

1942 1956 1968 1983 1985 2012 2016 Times

Figure 1: The brief history of artificial intelligence.

Table 1
Summarize properties for the symbolic systems and neural systems separately.
Systems Processing Methods Knowledge representation Primary algorithms Advantages Disadvantages
Strong generalization ability Weak at handling unstructured data
Symbolic systems Deductive reasoning Logical representation Logical deduction Good interpretability Weak robustness
Knowledge-driven Slow reasoning
Strong at handling unstructured data Weak generalizability (adaptability)
Neural systems
Inductive learning Distributed representation BP algorithms Strong robustness Lack of interpretability
(Sub-symbolic systems)
Fast learning Data-driven

provide explicit calculation processes. In contrast, symbolic systems offer enhanced appeal in terms of reasoning
and interpretability. For example, through deductive reasoning and automatic theorem proving, symbolic systems can
generate additional information and elucidate the reasoning process employed by the model.
Consequently, an increasing number of researchers have directed their attention towards the fusion of neural
systems and symbolic systems, aiming to achieve the third wave of AI: neural-symbolic learning systems [81, 80,
29, 85, 152, 35]. In a special NeurIPS 2019 lecture, Turing Award laureate Yoshua Bengio drew inspiration from Dr.
Daniel Kahneman’s renowned book "Thinking Fast and Slow" [56] to emphasize the need for a system-1-to-system-2
transformation in deep learning. Here, system 1 represents the intuitive, rapid, unconscious, nonlinguistic, and habitual
aspects, while system 2 embodies the deliberative, logical, sequential, conscious, linguistic, algorithmic, planning-
related, and reasoning-related facets. Indeed, since the 1990s, numerous researchers in the fields of artificial intelligence
and cognitive science have explicitly proposed the concept of dual processes that correspond to these contrasting
systems [50]. These highlight the necessity of combining neural systems and symbolic systems. By unifying these
two system types within a comprehensive framework, neural-symbolic learning systems can be created, endowing
AI with the capability to perform both perception and reasoning tasks. It is worth noting that the idea of integrating
neural systems and symbolic systems, referred to as hybrid connectionist-symbolic models, was initially introduced in
the 1990s [120].
Neural-symbolic learning systems leverage the combine strengths of both neural systems and symbolic systems
[79, 155, 99, 38, 126, 8, 44, 60, 118, 42, 31, 59, 66, 69]. To provide a comprehensive understanding, the survey
initially outlines key characteristics of symbolic systems and neural systems (refer to Table 1), including processing
methods, knowledge representation, etc. Analysis of Table 1 reveals that symbolic systems and neural systems exhibit
complementary features across various aspects. For instance, symbolic systems may possess limited robustness,
whereas neural systems demonstrate robustness. Consequently, neural-symbolic learning systems emerge as a means
to compensate for the shortcomings inherent in individual systems.
Moreover, we conduct an analysis of neural-symbolic learning systems from three key perspectives: efficiency,
generalization, and interpretability. As depicted in Figure 2, neural-symbolic learning systems excel in these areas.
Firstly, in terms of efficiency, neural-symbolic learning systems can reason quickly compared to pure symbolic systems,
thereby reducing computational complexity [150]. This accelerated computation can be attributed to the integration of

Dongran Yu et al.: Preprint submitted to Elsevier Page 2 of 29

 A Survey on Neural-symbolic Learning Systems

Neural systems

Neural-symbolic
systems

Efficiency

Symbolic systems

Low
Interpretability High
Figure 2: The advantages of neural-symbolic learning systems with respect to model efficiency, generalization, and
interpretability. The neural systems are black-box systems, while symbolic systems are white-box systems.

neural networks, as outlined in Section 2 (Learning for reasoning). Traditional symbolic approaches typically employ
search algorithms to navigate solution spaces, leading to increased computational complexity as the search space
grows in size. Secondly, with regard to generalization, neural-symbolic learning systems outperform standalone neural
systems in terms of their capacity for generalization. The incorporation of symbolic knowledge as valuable training
data enhances the model’s generalization abilities [58] (see Section 2 Reasoning for learning). Thirdly, in terms of
interpretability, neural-symbolic learning systems represent gray-box systems, in contrast to standalone neural systems.
By leveraging symbolic knowledge, these systems can provide explicit computation processes, such as traced reasoning
processes or chains of evidence for results [146]. Consequently, neural-symbolic learning systems have emerged as
vital components of explainable AI, yielding superior performance across diverse domains, including computer vision
[140, 28, 134, 58, 153, 14, 70], and natural language processing [72, 147, 125], etc.
Challenge: Symbolic systems and neural systems diverge in terms of their data representations and problem-
solving approaches. Symbolic systems rely on discrete symbolic representations and traditional search algorithms
to discover solutions, while neural systems employ continuous feature vector representations and neural cells to
learn mapping functions. Consequently, a significant challenge lies in designing a unified framework that seamlessly
integrates both symbolic and neural components. The aim is to strike a balance and select an appropriate combination
of symbolic and neural systems that aligns with the requirements of the specific problem [50].
To provide new readers with a comprehensive understanding of neural-symbolic learning systems, this paper
surveys representative research and applications of these systems. To summarize and systematically review related
works, there are several surveys conducted during the past few years [5, 43, 8, 12, 128, 65, 84, 132, 127, 41, 120, 119],
[61, 82]. For example, [5, 127] center around knowledge extraction techniques, which aligns with the first category
discussed in Section 2. On the other hand, [65, 12, 132, 84] provide detailed reviews from specific perspectives, such
as graph neural networks (GNNs), prior knowledge integration, explainable artificial intelligence (XAI), and statistical
relational learning. While surveys [8, 41] also cover neural-symbolic learning systems comprehensively, their focus
remains primarily theoretical, lacking a thorough introduction to specific techniques and related works. Therefore, an
urgent need arises to provide a comprehensive survey that encompasses popular methods and specific techniques (e.g.,
model frameworks, execution processes) to expedite advancements in the neural-symbolic field. Distinguishing itself
from the aforementioned surveys, this paper emphasizes classifications, techniques, and applications within the domain
of neural-symbolic learning systems.
Motivation: For our part, we do not seek to replace the above literature but complement them by offering a
comprehensive overview of the broader domain of neural-symbolic learning systems. This encompasses various
technologies, cutting-edge developments, and diverse application areas within the realm of neural-symbolic learning

Dongran Yu et al.: Preprint submitted to Elsevier Page 3 of 29

 A Survey on Neural-symbolic Learning Systems

systems. Additionally, this article caters to individuals engaged in the pursuit of integrating symbolic systems and
neural systems. With an emphasis on integration, we present a novel classification framework for neural-symbolic
learning systems in Section 2.
Our contributions can be summarized as follows:
1) We propose a novel taxonomy of neural-symbolic learning systems. Neural-symbolic learning systems are
categorized into three groups: learning for reasoning, reasoning for learning, and learning-reasoning.
2) We provide a comprehensive overview of neural-symbolic techniques, along with types and representations of
symbols such as logic knowledge and knowledge graphs. For each taxonomy, we provide detailed descriptions of the
representative methods, summarize the corresponding characteristics, and give a new understanding of neural-symbolic
learning systems.
3) We discuss the applications of neural-symbolic learning systems and propose four potential future research
directions, thus paving the way for further advancements and exploration in this field.
The remainder of this survey is organized as follows. In Section 2, we categorize the different methods of neural-
symbolic learning systems. Section 3 introduces the main technologies of neural-symbolic learning systems. We
summarize the main applications of neural-symbolic learning systems in Section 4. Section 5 discusses the future
research directions, after which Section 6 concludes this survey.

2. Categorization and Frameworks

In this section, we first introduce the theory and formal definition of neural-symbolic learning systems. We will
then proceed to summarize the taxonomy of neural-symbolic methods and discuss the frameworks that underpin these
taxonomies. In this survey, we use a color scheme to represent the different components of neural-symbolic learning
systems: green represents symbolic systems, and blue represents neural systems within the frameworks.
Neural-symbolic learning systems are hybrid models that leverage the strengths of both neural systems and
symbolic systems. To illustrate the relationships and characteristics between these systems, we provide a schematic
diagram in Figure 3. The green rectangle represents symbolic systems, which employ reasoning-based approaches
to find solutions. Symbolic systems typically operate on structured data, such as logic rules, knowledge graphs, or
time series data. Their fundamental unit of information processing is symbols. Through training, symbolic systems
acquire the solution space of a search algorithm for a specific task and output higher-level reasoning results. On the
other hand, the blue rectangle in the figure represents neural systems, which excel at learning-based approaches to
approximate the ground truth. Neural systems usually operate on unstructured data, such as images, videos, or texts,
and their primary information processing unit is a vector. Through training, neural systems learn a mapping function for
a specific task and output lower-level learning results. The outer blue box represents neural-symbolic learning systems,
which encompass the characteristics of both symbolic and neural systems (as summarized in Table 1). These systems
combine the reasoning capabilities of symbolic systems with the learning capabilities of neural systems to achieve a
comprehensive and integrated approach to problem-solving.
Based on the above description, the final objective of neural-symbolic learning systems is to find a function 𝐹
that can effectively map data 𝑥 and symbol 𝑠 (pre-defined or obtained via computation) to ground-truth 𝑦. The formal
definition is as follows:

∀(𝑥, 𝑦) ∈ 𝐷 𝐹 (𝑥, 𝑠) → 𝑦. (1)

In this survey, the neural systems mainly refer to deep learning (or deep neural networks), while symbolic systems
include symbolic knowledge [123, 22, 10] and symbolic reasoning techniques [18, 110], etc. The methodology of our
classification is determined by the integration mode between neural systems and symbolic systems, which has three
main integration methodologies. These three classifications are similar to [61] from essence but include it.

2.1. Learning for reasoning

In the first category, which we refer to as learning for reasoning, the goal is to leverage symbolic systems for
reasoning while incorporating the advantages of neural networks to facilitate finding solutions [103, 150, 35, 146,
30, 57, 33, 102, 95, 1, 113, 40, 107, 133, 92, 21, 141, 89, 20, 124, 52, 53, 46, 39]. In the context of learning for
reasoning, there are two main aspects to consider. The first aspect involves the use of neural networks to reduce the
search space of symbolic systems, thereby accelerating computation [103, 150, 102, 95, 1]. This can be achieved by
replacing traditional symbolic reasoning algorithms with neural networks. The neural network effectively reduces the

Dongran Yu et al.: Preprint submitted to Elsevier Page 4 of 29

 A Survey on Neural-symbolic Learning Systems

Table 2
Main Approaches of neural-symbolic learning systems. Taxonomies of the approaches and combination modes are presented
in Section 2. Methodical details can be found in Section 3. The symbols and neural networks used herein are introduced
in Appendix A, and "Agnostic" means arbitrary neural networks. "Serialization", "parallelization" and "interaction" are
combination modes, respectively. Applications are discussed in Section 5.

Representative works Taxonomies Methods Symbols Neural networks Applications

pLogicNet[103]
GNN Knowledge graph reasoning
ExpressGNN[150]

RNM[83] Learning for reasoning Accelerating CNN Classification

NMLN[86] (Serialization) GNN Knowledge graph reasoning

First-order logic
Classification and
NLIL[146] CNN, GNN
knowledge graph reasoning

NS-CL[81] Absracting RNN Visual question answering

HDNN[51] Agnostic Object recognition

SBR[25] CNN
Classification
SL[142] Agnostic
Propositional logic
LENSR[140] Visual relationship detection
Reguariziting
CA-ZSL[77] Reasoning for learning CNN,GNN

LSFSL[70] (Parallelization)

SEKB-ZSL[134] Knowledge graph GNN Few-shot classification

DGP[58] Transfering
CNN,GNN
KGTN[14]

PROLONETS[115] Propositional logic GNN Reinforcement learning

DeepProLog[80]
Agnostic
ABL[152]
Learning-reasoning Complex reasoning
GABL[11] Interacting First-order logic CNN
(Interaction)
WS-NeSyL[125] RNN

BPGR[148] CNN Visual relationship detection

search space, making the computation more efficient. The second aspect of learning for reasoning is the abstraction
or extraction of symbols from data using neural networks to facilitate symbolic reasoning [113, 40, 107]. In this
case, neural networks serve as a means of acquiring knowledge for symbolic reasoning tasks. They learn to extract
meaningful symbols from input data and use them for subsequent reasoning processes. The basic framework for

Dongran Yu et al.: Preprint submitted to Elsevier Page 5 of 29

 A Survey on Neural-symbolic Learning Systems

Data bases
Logic rules Symbol Symbolic Reasoning
representation
𝑦
Knowledge graphs systems
…
𝑠
Images Vector Neural
Videos representation systems Learning 𝑥
Texts
Solution
…
𝑥 System integration
Figure 3: Schematic diagram of two systems integration. Symbolic systems generally use reasoning technologies, such as
logic programs and search algorithms, to obtain a solution based on domain knowledge, such as first-order logic, knowledge
graphs, etc. The goal of neural systems is to learn a function from the training samples to predict a solution.

learning for reasoning is illustrated in Figure 4. As depicted in the figure, this type of model is characterized by
a serialization process, where the neural network component and the symbolic reasoning component are connected
sequentially. The neural network extracts relevant features or symbols from the input data, which are then used by the
symbolic reasoning module to perform higher-level reasoning tasks.

Neural network

accelerating 𝑦
Data bases
Logic rules Symbolic reasoning
Knowledge graphs
… 𝑥
𝑥 Solution
(a) First aspect

𝑦
Images
transforming
Videos Neural network Symbolic reasoning
Texts
…
𝑥 𝑥
(b) Second aspect Solution
Figure 4: Principle of Learning for reasoning. Its goal is to introduce neural networks to reasoning, a problem that it solves
primarily through reasoning technologies.

2.2. Reasoning for learning

In the second category, referred to as reasoning for learning, symbolic systems are utilized to support the learning
process of neural systems [51, 25, 142, 140, 77, 28, 91, 15, 6, 70, 74, 134, 58, 14, 28, 87, 36, 153, 96, 129, 101,
121, 97, 144, 54]. The underlying idea in reasoning for learning is to leverage neural systems for machine learning
tasks while incorporating symbolic knowledge into the training process to enhance performance and interpretability.
Symbolic knowledge is typically encoded in a format suitable for neural networks and used to guide or constrain the
learning process [51, 25, 142, 140]. For instance, symbolic knowledge may be represented as a regularization term in
the loss function of a specific task. This integration of symbolic knowledge helps improve the learning process and can
lead to better generalization and interpretability of the neural models. The basic principle of the reasoning for learning

Dongran Yu et al.: Preprint submitted to Elsevier Page 6 of 29

 A Survey on Neural-symbolic Learning Systems

approach is illustrated in Figure 5. This type of model is characterized by parallelization, where the neural system and
symbolic system operate in parallel during the learning process. The neural network component learns from the data,
while the symbolic system provides additional knowledge or constraints to guide the learning process.

Data bases
Logic rules Symbolic reasoning
Knowledge graphs
𝑦
…
Constraining
Images
Videos
Texts Neural network
…
𝑥 Solution
𝑥

Figure 5: Principle of reasoning for learning. It introduces symbolic knowledge to neural networks, and the main body relies
on neural networks to get solutions.

2.3. Learning-reasoning
In the third category, referred to as learning-reasoning, the interaction between neural systems and symbolic
systems is bidirectional, with both paradigms playing equal roles and working together in a mutually beneficial way
[80, 152, 11, 125, 148, 47]. The goal of learning-reasoning is to strike a balance between the involvement of neural
systems and symbolic systems in the problem-solving process. In this approach, the output of the neural network
becomes an input to the symbolic reasoning component, and the output of the symbolic reasoning becomes an input
to the neural network. By allowing the neural systems and symbolic systems to exchange information and influence
each other iteratively, this approach aims to leverage the strengths of both paradigms and enhance the overall problem-
solving capability. For example, incorporating symbolic reasoning techniques like abduction enables the design of
connections between deep neural networks and symbolic reasoning frameworks [79, 152, 148]. In this case, the neural
network component generates hypotheses or predictions, which are then used by the symbolic reasoning component
to perform logical reasoning or inference. The results from symbolic reasoning can subsequently be fed back to the
neural network to refine and improve the predictions. The basic principle of learning-reasoning is illustrated in Figure
6, where the interaction between neural systems and symbolic systems occurs in an alternating fashion. This mode of
combining both technologies allow for iterative learning and reasoning, enabling a deeper integration of neural and
symbolic approaches. By embracing bidirectional interaction and iterative exchange of information between neural
systems and symbolic systems, learning-reasoning approaches aim to maximize the strengths of both paradigms and
achieve enhanced problem-solving capabilities in various domains.

𝑦
Images
transforming
Videos Symbolic reasoning
Neural network
Texts
…
𝑥 constraining 𝑥
Solution
Figure 6: Principle of learning-reasoning. It combines neural networks and reasoning technologies as an alternate process,
and both together to output a solution.

In summary, the above three taxonomies are heterogeneous multi-module architectures in [119]. Learning for
reasoning and reasoning for learning are loosely coupled while learning-reasoning is tightly coupled. According to
the taxonomy of the neural-symbolic method presented in this paper, we summarize the existing main approaches from
six dimensions in Table 2: representative works, taxonomies, methods, symbols, neural networks and applications.
In the Section 3, we will introduce the details of these approaches, discussing their methodologies, techniques,
and characteristics. This will provide readers with a deeper understanding of how neural and symbolic systems are
combined in various ways to tackle different problems.

Dongran Yu et al.: Preprint submitted to Elsevier Page 7 of 29

 A Survey on Neural-symbolic Learning Systems

Table 3
Main characteristics of the selected methods.
Approaches Inputs Technology Tools Mechanism/Objective
pLogicNet [103] x,s SRL MLN learn a joint probability distribution
ExpressGNN [150] x,s SRL MLN learn a joint probability distribution
NLIL[146] x ILP Transformer use a Transform to learn rules based ILP
reason based on parsing symbols
NS-CL[81] x quasi-symbolic program concept parser
for images and questions
HDNN[51] x,s regularization t-norm learn a student network based on knowledge
learn a model with logic knowledge
SBR[25] x,s regularization t-norm
as a constraint of the hypothesis space
design a semantic loss to
SL[142] x,s regularization arithmetic circuits
act as a regularization term
align distributions between deep learning
LENSR[140] x,s regularization d-DNNF
and propositional logic
CA-ZSL[77] x,s regularization GCN learn a conditional random field
learn a deep learning model
SEKB-ZSR[134] x,s knowledge transfer GCN
with powerful generalization
DGP[58] x,s knowledge transfer GCN learn network embedding with semantics
KGTN [14] x,s knowledge transfer GGNN transfer semantic knowledge into weights
transform knowledge into
PROLONETS [115] x,s knowledge transfer decision tree
neural network parameters
construct an interface between
DeepProLog [80] x,s ProbLog SDD
probLog program and the deep learning models
minimize inconsistency between
ABL[152] x,s abductive reasoning SLD
pseudo-labels and symbolic knowledge
learn an encoder-decoder
WS-NeSyL[125] x,s ProLog SDD
constrained by logic rules
learn a model that fits both
BPGR[148] x,s SRL MLN
the ground truth and FOL

3. Methods of neural-symbolic learning systems

This section introduces the methods used in neural-symbolic learning systems in three main categories. We aim
to distill the representative ideas that provide evidence for the integration between neural networks and symbolic
systems, identify the similarities and differences between different methods, and offer guidelines for researchers. The
main characteristics of these representative methods are summarized in Table 3.

3.1. Learning for reasoning

Learning for reasoning methods leverage neural networks to accelerate the search speed of symbolic reasoning, or
to abstract unstructured data for symbolic reasoning. To accelerate symbolic reasoning, we first introduce approaches
based on SRL [64], such as pLogicNet [103] and ExpressGNN [150]. These approaches use neural networks to
parameterize the posterior computation of the probabilistic graphical models, which accelerates the search process
on solution space. Meanwhile, we investigate several methods based on ILP, such as NLIL [146], in which NLIL
automatically induces new logic rules from the data for model learning and reasoning. Next, for abstracting unstructured
data, we introduce several representative approaches such as NS-CL[81]. Specific details about the different models
are presented below.

Dongran Yu et al.: Preprint submitted to Elsevier Page 8 of 29

 A Survey on Neural-symbolic Learning Systems

3.1.1. Accelerating symbolic reasoning

In the case of Markov logic networks (MLNs), logic rules are encoded into an undirected graph, and graphical
models’ inference techniques are used to solve problems. However, performing computations on large-scale graphical
models can be challenging. On the other hand, neural networks can easily scale to large datasets, but they cannot directly
integrate logic knowledge. While several approximate inference methods have been proposed to address this issue,
they often come with high computational costs [116, 90, 117, 100, 62, 6]. To overcome these challenges, researchers
have introduced two models: Probabilistic Logic Neural Networks (pLogicNet) [103] and ExpressGNN [150]. Both
of these models aim to address the triplet completion problem in knowledge graphs by treating it as an inference
problem involving hidden variables in a probability graph. They achieve this by combining variational expectation-
maximization (EM) and neural networks to approximate the inference process. The basic process can be summarized
as follows:
(1) Model Construction: Both pLogicNet and ExpressGNN start by constructing a joint probabilistic model that
combines the logic rules encoded in the knowledge graph with the neural network model. This joint model captures the
dependencies between observed and hidden variables. (2) Inference Approximation: To perform inference, pLogicNet
and ExpressGNN utilize a variational EM approach. They define an approximate posterior distribution over the hidden
variables and iteratively optimize the model parameters and approximate posterior distribution to maximize the joint
probability. (2) Neural Network Learning: During the optimization process, the neural network parameters are updated
based on the observed data, using techniques such as backpropagation and gradient descent. This allows the neural
network to learn the underlying patterns and relationships in the data. (3) Inference and Prediction: After training, the
models can be used for inference and prediction tasks. They can estimate the probabilities of unobserved variables,
complete missing values, or make predictions based on the learned dependencies.
The main goal of pLogicNet and ExpressGNN is to combine the strengths of logic-based reasoning and neural
networks in a unified framework. By leveraging neural networks’ scalability and the expressiveness of logic rules,
these models offer an solution for reasoning tasks in knowledge graphs. ExpressGNN improves the inference network
of pLogicNet by using a graph neural network (GNN) to replace the flattened embedding table and adding a tunable
part to the entity embedding to alleviate the problem of isomorphic nodes having the same embedding. The specific
framework of ExpressGNN is illustrated in Figure 7.

GNN: Feature representation Boolean random variable

MLN：Factor graph
（subject, predicate, object） r e2)
1(e1,
（e1, r1, e2 ） … …
（e2, r1, e3 ） r1(e2, e3)
e2 （e1, r3, e3 ） Rule sets
r1 （e1, r2, e4 ） r3(e1, e3)
e4 e1
r2 r3 … … … …
r1 （e7, r8, e10） r1(e1, e2) r1(e2, e3) r3(e1, e3) rk(em, en)
e3
（em, rk, en） rk(em, en)

Variational EM
Symbol space
Continuous space

Figure 7: The framework of the ExpressGNN model. The model contains continuous space and symbolic space, and the
variational EM algorithm acts as a bridge that connects the continuous space and symbolic space. Note, in the knowledge
graph, 𝑒𝑘 is the entity and 𝑟𝑘 is the relation.

Indeed, relying solely on manually constructed logic rules may not capture the complete knowledge present in the
data. To address this limitation, researchers have explored approaches to automatically learn logic rules from data. Two
notable methods in this regard are the extensions of Markov logic networks (MLNs) proposed by Marra et al. [86, 83]
and differentiable inductive logic programming (ILP) models. Marra et al. extended MLNs by designing a general
neural network architecture that can automatically learn the potential functions of MLNs from the original data. By
training the neural network on labeled data, the model learns to capture the underlying patterns and relationships in
the data, effectively learning logic rules that approximate the true structure of the domain. This approach enables the
integration of neural networks and symbolic reasoning, allowing the model to learn logic rules directly from the data.
Differentiable ILP is another approach that combines neural networks and logic to learn rules . It extends traditional
ILP methods [67] by introducing differentiable operations that enable the incorporation of neural networks into the

Dongran Yu et al.: Preprint submitted to Elsevier Page 9 of 29

 A Survey on Neural-symbolic Learning Systems

ILP framework. This allows the model to learn logic rules by leveraging the expressive power of neural networks.
Several differentiable ILP models have been proposed [37, 35, 13, 109, 98], including 𝜕ILP [35], which uses predefined
templates to construct logic rules and applies forward reasoning for inference. 𝜕ILP is capable of learning effective
logic rules even in the presence of noisy data, making it robust to imperfect input.
The current methods for learning logic rules often face limitations in expressiveness and computational feasibility.
Approaches such as expressing the chain rule as a Horn clause and controlling the search length, number of
relationships, and entities have been employed to address these challenges [21, 145, 45]. However, the limited
expressive power of these complex logic rules can hinder their effectiveness. To overcome these limitations, Yang
et al. introduced neural logic inductive learning (NLIL) [146]. NLIL is a differentiable ILP model that extends the
multi-hop reasoning framework to address general ILP problems. It allows for the learning of complex logic rules,
including tree and conjunctive rules, which offer greater expressiveness compared to traditional approaches. NLIL
leverages neural networks to learn logic rules from data and provides explanations for patterns observed in the data.
It first converts a logical predicate into a predicate operation, and then transforms all the intermediate variables
into predicate operation representations of the head and tail entities, and such the head and the tail variables can be
represented by randomly initialized vectors in the concrete implementation, so as to get rid of data dependency; then
such a predicate operation forms the atom of the logical paradigm, which greatly expands the expression capability
of logical predicates such as from chains to trees. Next, the NLIL model further extends the expressive power of the
generated logical paradigm by combining atoms using logical connectives (and, or, not). Indeed, the NLIL model uses a
hierarchical transformer model 1 to efficiently compute the intermediate parameters to be learned, including the vectors
of logical predicates and the corresponding parameters of the attention mechanism (the weighted summation).
In NLIL, logic rules are grounded through matrix multiplication. For example, consider the logic rule 𝐹 𝑟𝑖𝑒𝑛𝑑𝑠(𝑥, 𝑦) ⇒
𝑆𝑚𝑜𝑘𝑒𝑠(𝑥), where constants 𝐶 = {𝐴, 𝐵} are one-hot vectors, such as 𝑉𝐴 and 𝑉𝐵 , and predicates 𝐹 𝑟𝑖𝑒𝑛𝑑𝑠 and 𝑆𝑚𝑜𝑘𝑒𝑠
are mapped to matrices such as 𝑀𝐹 𝑟𝑖𝑒𝑛𝑑 (𝐴, 𝐵) is a score of 𝐴 and 𝐵 that are related by 𝐹 𝑟𝑖𝑒𝑛𝑑. The score of grounding
is 𝑉𝐵 = 𝑉𝐴 𝑀𝐹 𝑟𝑖𝑒𝑛𝑑 .

3.1.2. Abstracting unstructured data into symbols

Mao et al. [81] proposed the Neuro-Symbolic Concept Learner (NS-CL), which uses neural symbolic reasoning as
a bridge to jointly learn visual concepts, words, and the semantic parsing of sentences without the need for the explicit
supervision of any of them. NS-CL builds an object-based scene representation and translates sentences into symbolic
programs.
NS-CL is designed to solve tasks in visual question answering (VQA). This model includes three modules: a visual
perception module, semantic parsing module, and symbolic reasoning module. The visual perception module extracts
object-based symbolic representation for a scene in an image. The semantic parsing module transforms a question into
an executable program. Finally, the symbolic reasoning module applies a quasi-symbolic program executor to infer the
answer based on the representation and an executable program.
An example is shown in Figure 8. Given the input image and question, the visual perception module uses a
pre-trained mask R-CNN [26] to generate object features for all objects, then applies a similarity-based metric to
classify objects. The semantic parsing module translates a natural language question into an executable program with
a hierarchy of primitive operations, represented in a domain-specific language (DSL) designed for VQA. The symbolic
reasoning module is a collection of deterministic functional modules designed to realize all logic operations specified
in the DSL. It takes object concepts and a program as input to derive the answer. The optimization objective of NS-CL
is composed of both a visual perception module and a semantic parsing module.
Conclusion: Based on the discussion of related works in the category of learning for reasoning, we can summarize
two principles from the perspective of neural networks: (1) Accelerator. Symbolic reasoning often involves searching
for solutions in a solution space, which can be computationally expensive. In learning for reasoning approaches, neural
networks are used as accelerators to speed up the search process. By leveraging the learning capabilities of neural
networks, the search space can be reduced or optimized, leading to faster and more efficient reasoning. Techniques
such as reinforcement learning can be employed to guide the search process and improve the overall performance
of symbolic systems. (2) Transformer. Symbolic systems traditionally operate on structured data represented by
symbols. However, neural networks excel at processing unstructured data such as images, texts, and videos. In
learning for reasoning approaches, neural networks play a crucial role in transforming unstructured data into symbolic
representations that can be processed by symbolic reasoning systems. By abstracting and extracting relevant features
1 Transformer is a tool that learns the relation between sequence data. Here, learning rules is a process of sequence calculation.

Dongran Yu et al.: Preprint submitted to Elsevier Page 10 of 29

 A Survey on Neural-symbolic Learning Systems

obj1
obj2
obj3
obj4

visual perception module symbolic

Answer: Cylinder
reasoning
module

Q: What is the shape of the red object? semantic parsing module

Query(Shape, Filter(Red))
domain-specific language

Figure 8: The framework of the NS-CL. The perception module begins by parsing visual scenes into object-based deep
representations, while the semantic parser parse sentences into executable programs. A symbolic reasoning process bridges
two modules.

from the unstructured data, neural networks provide the symbolic systems with meaningful inputs for reasoning and
decision-making.

3.2. Reasoning for learning

Reasoning for learning methods can be broadly divided into regularization models and knowledge transfer models.
Regularization models: Regularization models integrate symbolic knowledge into the training process by adding
regular terms to the objective function of the model. These regular terms serve as constraints or penalties that encourage
the model to adhere to the symbolic knowledge during training. By incorporating prior knowledge, regularization
models aim to improve the performance and interpretability of the trained model. Different regularization models may
employ different strategies for modeling symbolic knowledge as regular terms, such as semantic embeddings, logical
rules, or semantic constraints. Examples of regularization models include [51, 25, 142, 140, 77, 28, 91, 15, 70, 74].
Knowledge transfer models: Knowledge transfer models establish connections between different domains or spaces,
such as the visual space and the semantic space, and transfer symbolic knowledge from one domain to another to
support the learning process. These models leverage existing symbolic knowledge or semantic information to guide
the learning of models in a different domain. For example, semantic knowledge can be transferred from the semantic
space to the visual space to aid in tasks such as zero-shot learning or semantic segmentation. Knowledge transfer models
enable the utilization of symbolic knowledge in a different context to improve the performance and generalization of
the model. Examples of knowledge transfer models include [134, 58, 14].

3.2.1. Regularization models

Indeed, Hu et al. [51] introduced the HDNN framework, which leverages the concept of knowledge distillation
to harness the power of logic rules in training deep neural networks. The framework consists of two components: a
teacher network and a student network. In HDNN, the teacher network encodes logic rules and guides the student
network during training. That is to say, the teacher network learns information from the labeled data and logic rules
(unlabeled data) and teaches the student network by the loss function. Based on the above process, the structured
information encoded by logic rules can constrain the learning of the student network.
Specifically, the rule knowledge distillation module includes the student network 𝑝𝜃 (𝑦|𝑥) and the teacher network
𝑞(𝑦|𝑥). The model needs to meet the following conditions: 1) The probability distribution 𝑝𝜃 (𝑦|𝑥) of the student
network should be as close as possible to the probability distribution 𝑞(𝑦|𝑥) of the teacher network; 2) The teacher
network should obey the logic rules to the greatest extent possible. The formal representation of the objective and two
conditions is shown in Eq. (2) and Eq. (3).

Dongran Yu et al.: Preprint submitted to Elsevier Page 11 of 29

 A Survey on Neural-symbolic Learning Systems

∑
𝑁
𝜃 (𝑡+1) = arg min𝜃∈Θ 1∕𝑁 (1 − 𝜋)𝑙(𝑦𝑛 , 𝜎𝜃 (𝑥𝑛 )) + 𝜋𝑙(𝑠𝑛 (𝑡) , 𝜎𝜃 (𝑥𝑛 )), (2)
𝑛=1

⎧min ∑
𝑞,𝜉≥0 𝐾𝐿(𝑞(𝑌 |𝑋)‖𝑝𝜃 (𝑌 |𝑋)) + 𝐶 𝑙,𝑔𝑙 𝜉𝑙,𝑔𝑙
⎪
⎨𝜆𝑙 (1 − 𝐸𝑞 [𝑟𝑙,𝑔𝑙 (𝑋, 𝑌 )]) ≤ 𝜉𝑙,𝑔𝑙 (3)
⎪𝑔𝑙 = 1, ..., 𝐺 , 𝑙 = 1, ..., 𝐿
⎩ 𝑙

In Eq. (2), 𝜋 is the limitation parameter used to calibrate the relative importance of the two objectives; 𝑥𝑛 represents
the training data, while 𝑦𝑛 the label of the training data; 𝑙 denotes the loss function selected according to specific
applications (e.g., the cross-entropy loss for classification); 𝑠(𝑡)
𝑛 is the soft prediction vector of 𝑞(𝑦|𝑥) on 𝑥𝑛 at iteration
𝑡; 𝜎𝜃 (𝑥) represents the output of 𝑝𝜃 (𝑦|𝑥); the first term is the student network, and the second term is the teacher
network. In Eq. (3), 𝜉𝑙,𝑔𝑙 ≥ 0 is the slack variable for the respective logic constraint; 𝐶 is the regularization parameter;
𝑙 is the index of the rule; 𝑔𝑙 is the index of the ground rule; 𝜆𝑙 is the weight of the rule.
Different from the knowledge distillation framework, certain approaches incorporate logical knowledge as a
constraint within the hypothesis space. These methods involve encoding a logic formula, either propositional or
first-order, into a real-valued function that serves as a regularization term for the neural model. An example of
such an approach is semantic-based regularization (SBR) proposed by Diligenti et al. [25]. SBR combines the
strengths of classic machine learning, with its ability to learn continuous feature representations, and symbolic
reasoning techniques, with their advanced semantic knowledge reasoning capabilities. SBR is applied to address
various problems, including multi-task optimization and classification. Following the classical penalty approach for
constrained optimization, constraint satisfaction can be enforced by adding a term that penalizes the violation of these
constraints into the loss of the model.
Building upon the idea of semantic-based regularization (SBR), Xu et al. [142] introduced a novel approach called
semantic loss (SL). SL combines the power of propositional logic reasoning with deep learning architectures by
incorporating the output of the neural network into the loss function as a constraint for the learnable network. This
enables the neural network to leverage the reasoning capabilities of propositional logic to improve its learning ability.
In contrast to SBR, SL takes a different approach to incorporating logic rules into the loss function. It encodes the
logic rules using an arithmetic circuit, specifically a Sentential Decision Diagram (SDD) [18], which allows for the
evaluation of the model. This encoding serves as an additional regularization term that can be directly integrated into
an existing loss function. The formulation of SL is provided in Equation (4).
∑ ∏ ∏
𝐿𝑠 (𝛼, 𝑝) ∝ − log 𝑝𝑖 (1 − 𝑝𝑖 ), (4)
𝑥⊧𝛼 𝑖∶𝑥⊧𝑋𝑖 𝑖∶𝑥⊧¬𝑋𝑖

where 𝑝 is a vector of probabilities from the prediction of the neural network, 𝛼 is a propositional logic, 𝑥 is the
instantiation of 𝑋𝑖 , and 𝑥 ⊧ 𝛼 is a state 𝑥 that satisfies a sentence 𝛼.
Notably, the aforementioned methods do not employ explicit knowledge representation techniques, leading to an
unclear computational process for symbolic knowledge. To address this issue, some researchers have chosen to utilize
tools that can model symbols, such as d-DNNF (as mentioned in Section A). Xie et al. [140] integrated propositional
logic into a relationship detection model and proposed a logic embedding network with semantic regularization
(LENSR) to enhance the relationship detection capabilities of deep models. The process of LENSR can be summarized
as follows: 1) The visual relationship detection model predicts the probability distribution of the relation predicate for
each image; 2) The prior propositional logic formula related to the sample image is expressed as a directed acyclic
graph by d-DNNF, after which GNN is used to learn its probability distribution; 3) An objective function is designed
that aligns the above two distributions.
Figure 9 presents a schematic diagram of LENSR, which uses a propositional logic of the form 𝑃 ⇒ 𝑄. In this
example, the predicate 𝑃 represents 𝑤𝑒𝑎𝑟(𝑝𝑒𝑟𝑠𝑜𝑛, 𝑔𝑙𝑎𝑠𝑠𝑒𝑠), and the predicate 𝑄 represents 𝑖𝑛(𝑔𝑙𝑎𝑠𝑠𝑒𝑠, 𝑝𝑒𝑟𝑠𝑜𝑛). The
ground truth of the input image and the corresponding propositional logic (prior knowledge) are on the left. The
directed acyclic graph of propositional logic d-DNNF is then sent to the Embedder 𝑞 (Embedder is a graph neural
network (GNN) [63] that learns the vector representation) to obtain 𝑞(𝐹𝑥 ), which is the embedding of the propositional

Dongran Yu et al.: Preprint submitted to Elsevier Page 12 of 29

 A Survey on Neural-symbolic Learning Systems

logic knowledge. On the right is the relation label predicted by the detection network. The predicted labels are then
combined into a conjunctive normal form ℎ(𝑥) = ∧𝑝𝑖 to construct a directed acyclic graph d-DNNF, which is sent to
the Embedder 𝑞 to obtain 𝑞(ℎ(𝑥)), the embedding of the predicted propositional logic. The optimization goal of LENSR
is shown in Eq. (5). Here, 𝐿𝑡𝑎𝑠𝑘 represents the loss of a specific task, 𝜆 is a hyperparameter that acts as a balance factor,
and 𝐿𝑙𝑜𝑔𝑖𝑐 is the loss of propositional logic (that is, the distance between vector 𝑞(𝐹 𝑥) and vector 𝑞(ℎ(𝑥))).

𝐿 = 𝐿𝑡𝑎𝑠𝑘 + 𝜆𝐿𝑙𝑜𝑔𝑖𝑐 , (5)

𝐿𝑙𝑜𝑔𝑖𝑐 = ‖𝑞(𝑓 ) − 𝑞(∧𝑝𝑖 )‖2 . (6)

Ground truth Predict result

𝑳𝒍𝒐𝒈𝒊𝒄 = ||𝒒 Fx − 𝒒(∧ 𝒑𝒊 )||𝟐𝟐

𝒒:
Embedder

p1 Λ p2 Λ p3
P => Q ^ ^
p1: wear (person, glasses)
P: wear (person, glasses)
⋁ p2:next-to(person, board)
Q: in (glasses, person)
-p q p1 p2 p3 p2:above(person, desk)
………….
Fx ∧𝑖 𝒑𝒊

Figure 9: The framework of the LENSR. The GCN-based embedder 𝑞 projects a logic graph to the vector space, satisfying
the requirement that the distribution of the projected result is as close to the distribution of the real label as possible. We
use this embedding space to form logic losses that regularize deep neural networks for a target task.

However, LENSR models a local dependency graph for each logic formula and only captures local knowledge
information, which may result in the poor expressive ability of the model. To solve this problem, the researcher started
to model a global dependency graph for all logic formulas, which can improve the expressive ability of the model and
effectively capture uncertainty [148].
The above methods utilize logic rules as prior knowledge. In contrast, Luo et al. [77] focused on knowledge graphs
and proposed a context-aware zero-shot recognition method (CA-ZSL) to address the zero-shot detection problem.
CA-ZSL constructs a model based on deep learning and conditional random fields (CRF) and leverages knowledge
graphs, which represent the semantic relationships between classes, to assist in identifying objects from unseen classes.
The framework of CA-ZSL is depicted in Figure 10. In this framework, individual and pairwise features are extracted
from the image. The instance-level zero-shot inference module utilizes individual features to generate a unary potential
function, while the relationship inference module employs pairwise features and knowledge graphs to generate a binary
potential function. Finally, based on the conditional random field constructed by these two potential functions, the label
of the unseen objects is predicted.
CA-ZSL incorporates a knowledge graph, which includes the GCN-encoded embedding, into the computation
of the binary potential function within the conditional random field, thereby facilitating the learning process of the
model. The objective of optimization is to maximize the joint probability distribution of the conditional random field,
as depicted in Equation (7). In the equation, 𝜃 denotes the unary potential function, 𝜓 represents the binary potential
function, 𝑐𝑖 represents the class, 𝐵𝑖 represents the object region in the image, 𝛾 is the balance factor, and 𝑁 denotes
the number of objects.
∑ ∑
𝑃 = (𝑐1 ...𝑐𝑁 |𝐵1 ...𝐵𝑁 ) ∝ 𝑒𝑥𝑝( 𝜃(𝑐𝑖 |𝐵𝑖 ) + 𝛾 𝜓(𝑐𝑖 , 𝑐𝑗 |𝐵𝑖 , 𝐵𝑗 )). (7)
𝑖 𝑖≠𝑗

Dongran Yu et al.: Preprint submitted to Elsevier Page 13 of 29

 A Survey on Neural-symbolic Learning Systems

Knowledge Graph

class1
Object detection model
class5

class6
Images
Figure 10: The framework of the CA-ZSL. The features of individual objects and pairwise features are extracted from
the image and input into an instance-level zero-shot inference module and a relationship inference module respectively. In
combination with the knowledge graph, the unary potential function and binary potential function of CRF are generated
respectively to predict the labels of objects.

3.2.2. Knowledge transfer models

Knowledge transfer integrates knowledge graphs (pre-defined) that represent semantic information into deep
learning models, which compensates for the lack of available data through the transfer of semantic knowledge. In
this survey, we mainly introduce representative approaches in image classification [135, 14] such as zero-shot learning
[134, 58], few-shot learning [151] [135, 14] and reinforcement learning [115] for knowledge transfer.
In the context of visual tasks, researchers have proposed several zero-shot recognition models that leverage
semantic representation and knowledge graphs to compensate for the limitations of visual data. Two notable models
in this regard are SEKB-ZSL (Zero-shot Recognition via Semantic Embeddings) [134] and DGP (Dense Graph
Propagation Module) [58]. These models employ the semantic classifier weights derived from the knowledge graph,
which encompasses both seen and unseen classes, to guide or supervise the learning of visual classifier weights, thus
facilitating knowledge transfer. The underlying principles can be summarized as follows: In the visual space, SEKB-
ZSL employs Convolutional Neural Networks (CNN) to extract visual features from images and learns a visual classifier
for the seen classes. In the semantic space, Graph Convolutional Networks (GCN) are utilized to learn node features
within the knowledge graph and derive a semantic classifier for class labels. Subsequently, the weights of the class
semantic classifier are employed to supervise the learning process of visual classifier weights, enabling the transfer of
semantic knowledge to novel classes.
DGP introduces improvements over SEKB-ZSL. Firstly, it addresses the over-smoothing issue in GCN by reducing
the number of graph convolution layers from six to two. Additionally, the attention mechanism is employed to calculate
connection weights between nodes in the knowledge graph, thereby enhancing the node connectivity. The principle of
DGP is depicted in Figure 11. The entire model is trained end-to-end, and the loss function is defined as the similarity
between the weights of the two module classifiers, as shown in Eq. (8). In the equation, 𝑀 represents the number of
classes, 𝑃 represents the dimension of the weight vectors, 𝑊𝑖,𝑗 denotes the weights of the visual classifier, and 𝑊 ′ 𝑖,𝑗
denotes the weights of the semantic classifier.

∑
𝑀 ∑
𝑃
2
𝐿 = 1∕2𝑀 (𝑊𝑖,𝑗 − 𝑊 ′ 𝑖,𝑗 ) . (8)
𝑖=1 𝑗=1

Transferring correlation information between classes can be beneficial for learning new concepts. DGP demon-
strates the importance of aligning the semantic classifier with the feature classifier. Building upon this idea, Chen et
al. [14] proposed the Knowledge Graph Transfer Network (KGTN) to address the few-shot classification problem. In
KGTN, knowledge graphs are utilized to capture and model the correlations between seen and unseen classes. These

Dongran Yu et al.: Preprint submitted to Elsevier Page 14 of 29

 A Survey on Neural-symbolic Learning Systems

CNN

𝑊
𝐿𝑜𝑠𝑠 = ||𝑊 − 𝑊 ′ ||2

GNN

Knowledge graph 𝑊′
Figure 11: The framework of the DGP. DGP is trained to predict the classifier weights 𝑊 for each node/class in a graph.
The weights for the training classes are extracted from the final layer of a pre-trained ResNet. The graph is constructed from
a knowledge graph, and each node is represented by a vector that encodes semantic class information (the word embedding
class in this paper). The network consists of two phases: a descendant phase (where each node receives knowledge from
its descendants) and an ancestor phase (where it receives knowledge from its ancestors).

knowledge graphs serve as a means to transfer knowledge and facilitate the learning process. The overall architecture
of KGTN is depicted in Figure 12.
Specifically, KGTN comprises three main parts: the feature extraction module, knowledge graph transfer module,
and prediction module. The feature extraction module uses CNN to extract the feature vector of images. The knowledge
graph transfer module uses a gated graph neural network (GGNN) to learn the knowledge graph node embedding. After
𝑇 iterations, the knowledge graph transfer module obtains the final weight 𝑤∗ , which has captured the correlation
between the seen and the unseen classes. The prediction module calculates the similarity between the weight 𝑤∗ and
the image feature to predict the probability distribution of the label.

CNN

x
Knowledge Graph

GGNN Similarity metric

𝐰∗

Figure 12: The framework of the KGTN model. It incorporates the prior knowledge of category correlation and makes
use of the interaction between category classifier weights, facilitating better learning of the classifier weights of unknown
categories.

Dongran Yu et al.: Preprint submitted to Elsevier Page 15 of 29

 A Survey on Neural-symbolic Learning Systems

In contrast to static domain knowledge, Silva et al. [115] introduced Propositional Logic Nets (PROLONETS),
which directly encode domain knowledge as a collection of propositional rules within a neural network. PROLONETS
not only incorporates domain knowledge into the model but also allows for the refinement of domain knowledge based
on the trained neural network. The framework of PROLONETS is illustrated in Figure 13. This approach enables the
neural network to leverage domain-specific information and improve its learning and reasoning capabilities.
PROLONETS aids in “warm starting" the learning process in deep reinforcement learning. The first step involves
knowledge representation, where policies and actions express domain knowledge in the form of propositional rules.
These rules are then encoded into a decision tree structure. The second step is neural network initialization, wherein the
nodes of the decision tree are directly transformed into neural network weights. This allows the agent to immediately
commence learning effective strategies in reinforcement learning. The final step is training, during which the initialized
network interacts with the environment, collecting data that is subsequently used to update parameters and rectify
domain knowledge.

position>-1

if position>-1 then left

if position<1 then left
position<1 angle<0 Environment
… state reward
if angle<0 then right angle<0 … right
Agent
Domain knowledge
left right

Decision tree Reinforcement learning

Figure 13: The framework of the PROLONETS. Domain knowledge is constructed into a decision tree that is then used
to directly initialize a PROLONET’s architecture and parameters; here, leaf nodes are actions and other nodes are policies
in reinforcement learning. The PROLONET can then begin reinforcement learning in the given domain, outgrowing its
original specification.

Let us consider the cart pole as an example. The state space of a cart pole is a four-dimensional vector: cart
position, cart velocity, pole angle, and pole velocity. The action space is a two-dimensional vector (left, right). Domain
knowledge can be expressed as "if the cart’s position is right of center, move left; otherwise, move right.". The decision
nodes of the tree become linear layers, leaves become action weights, and the final output is a sum of the leaves weighted
by path probabilities. Therefore, if 𝑝𝑜𝑠𝑖𝑡𝑖𝑜𝑛 > −1, the weight of the neural network is 𝑤 = {1, 0, 0, 0}, and the bias is
𝑏 = −1.
Conclusion: Based on the above text, we can summarize the following key factors in reasoning for learning. (1)
Knowledge representation. Symbolic knowledge is a kind of discrete representation. To achieve combination between
symbolic knowledge (discrete representation) and neural network (continuous representation), most methods usually
convert the symbolic knowledge into an intermediate representation, such as a graph, tree, etc. Moreover, another
approaches use fuzzy logic (such as t-norm) to assign soft truth degrees in the continuous set [0, 1]. In summary, it
may very well be among the essential representations grounded in the environment that form the foundation of a much
larger representational edifice that is needed for human-like general intelligence [50]. (2) Combining approaches.
One type of approach involves taking symbolic knowledge as a regularization term in the loss functions of the neural
networks. The others involve encoding symbolic knowledge into the structure of the neural networks as an initiative to
improve their performance. It is worth noting that logic rules usually are added as constraints to loss functions, while
knowledge graphs often enhance neural networks with information about relations between instances.

3.3. Learning-reasoning
In learning-reasoning approaches, learning and reasoning do not work in isolation but instead closely interact. This
is a development trend of neural-symbolic learning systems [80, 152, 11, 125, 148].
Based on ProbLog [23], Robin et al. [80] introduced neural facts and neural annotated disjunction (neural
AD) to propose a model that seamlessly integrates probability, logic, and deep learning, known as DeepProbLog.
DeepProbLog is a pioneering framework that combines a generic deep neural network with probabilistic logic in a
unique manner. It offers the advantage of enhanced expressive capability and enables end-to-end training for neural
networks and logical reasoning in a unified framework.

Dongran Yu et al.: Preprint submitted to Elsevier Page 16 of 29

 A Survey on Neural-symbolic Learning Systems

DeepProbLog is a probabilistic programming language that integrates deep learning through the use of "neural
predicates". These neural predicates serve as an interface between neural networks and symbolic reasoning. In
DeepProbLog, an image, for instance, is processed by a neural network, which outputs the distribution of each
class in the dataset as logical facts for symbolic reasoning. Specifically, neural networks are employed to process
simple concepts or unstructured data, generating inputs for symbolic reasoning in DeepProbLog. Symbolic reasoning
in DeepProbLog utilizes SDD (Sentential Decision Diagrams) [18] to construct a directed graph, which is then
transformed into an arithmetic circuit for inference and answering queries. To enable end-to-end training that bridges
continuous embedding and discrete symbols, DeepProbLog leverages the gradient semiring [32] as an optimization
tool. The framework of DeepProbLog is visually depicted in Figure 14.

Output： query Logical reasoning

× ×

Output 1 output 2 output 3 …… output n

Machine learning

Input：Unstructured data/ simple structured data

Figure 14: The framework of DeepProLog. Machine learning is responsible for mapping the input (unstructured data or
simple structured data) to the distribution of categories (if there are n categories, the output distribution is 1×n). Logical
reasoning is a complex problem described by ProbLog, which is constructed as an arithmetic circuit to solve the complex
problem. Here, the root node is the query, while the leaf nodes are neural predicates and other (non-neural network output)
probabilistic facts.

In a different approach to DeepProbLog, Zhou et al. [152] proposed abductive learning (ABL) as a framework
that combines abductive reasoning [2] with induction. Abductive reasoning, which is a form of logical reasoning,
involves inferring the best explanation for given observations or evidence. ABL leverages both induction, which
is a key component of modern machine learning, and abduction, which is the process of generating hypotheses or
explanations, in a mutually beneficial way. ABL provides a unified framework that bridges machine learning and logical
reasoning, allowing for the integration of both approaches to improve the overall learning and reasoning process. This
framework offers a new perspective and methodology for effectively combining machine learning techniques with
logical reasoning techniques.
In more detail, given raw data (This raw data only includes the data features and a label of true or false, with
no label of the class.), an initialized classifier, and a knowledge base (KB), the raw data is fed into the initialized
classifier to obtain pseudo-labels in machine learning. These pseudo-labels (pseudo-grounding) are then transformed
into symbolic representations that can be accepted by logical reasoning. Next, ABL uses ProLog as the KB and adopts
abductive reasoning technology to abduct the pseudo-labels and rules. That is to say, logical reasoning minimizes
the inconsistency between the symbolic representation and the KB to revise pseudo-labels, then outputs the deductive
labels. Finally, a new classifier is trained by the deductive labels and the raw data, which replaces the original classifier.
The above is an iterative process that continues until the classifier is no longer changed or the pseudo-labels are

Dongran Yu et al.: Preprint submitted to Elsevier Page 17 of 29

 A Survey on Neural-symbolic Learning Systems

consistent with the KB. ABL is a special kind of weakly supervised learning, in which the supervision information
comes not only from the ground-truth labels but also from knowledge abduction.
Based on the ABL framework, Tian et al. [125] proposed a weakly supervised neural symbolic learning model
(WS-NeSyL) for cognitive tasks with logical reasoning. The difference between WS-NeSyL and ABL is that ABL uses
a metric of minimal inconsistency in logical reasoning, while WS-NeSyL adopts sampling technology. In WS-NeSyL,
to provide supervised information for the reasoning process in complex reasoning tasks, the neural network is designed
as an encoder-decoder framework that includes an encoder and two decoders (perceptive decoder and cognitive
decoder). The encoder can encode input information as a vector, while the perceptive decoder decodes the vector to
predict labels (pseudo-labels). According to these pseudo-labels and the sampled logic rules from the knowledge base,
the cognitive decoder to reason results. To supervise the reasoning of the cognitive decoder, WS-NeSyL provides a
back search algorithm to sample logic rules from the knowledge base to act as labels that are used to revise the predicted
labels. To solve the sampling problem, WS-NeSyL introduces a regular term of logic rules. The whole model is trained
iteratively until convergence.
The knowledge base is an important factor in logical reasoning, and different knowledge bases are used by different
reasoning technologies. The above approaches use probabilistic logic programming language (ProbLog) as their
knowledge base; notably, they only consider that neural networks can provide facts for the knowledge base, and do
not quantify how many logic rules should be triggered by the neural networks. To resolve this issue, Yu et al. [148]
proposed a bi-level probabilistic graphical reasoning framework, called BPGR. To quantify the amount of symbolic
knowledge that is triggered, BPGR uses MLN to model all logic rules. For instance, MLN can express the time at
which a logic rule is true in the form of a potential function.
BPGR includes two parts: the visual reasoning module (VRM) and the symbolic reasoning module(SRM). VRM
extracts the features of objects in images and the inferred labels of objects and relationships. SRM uses symbolic
knowledge to guide the reasoning of VRM in a good direction, which acts as an error correction. In terms of the model
framework, more concretely, SRM is a double-layer probabilistic graph that contains two types of nodes: one is the
reasoning results of the VRD model in the high-level structure, and the other is the ground atoms of logic rules in
the low-level structure. When the probabilistic graphical model is constructed, BPGR can be efficiently trained in an
end-to-end manner by the variational EM algorithm. An overall framework of BPGR is provided in Figure 15.

update

Visual relationship detection

MLN

Visual reasoning module Symbolic reasoning module

Figure 15: The framework of the BPGR. This model is a two-layers probabilistic graphical model which consists of a visual
reasoning model and a symbolic reasoning module. Here, The high-level structure is the result of the visual reasoning
module, while the low-level structure is the ground atom of logic rules. The model is trained to output reasoning results of
the visual reasoning module based on symbolic knowledge. Note that G represents the grounding operator, the solid line
represents ground truth edges, and the dotted line represents pseudo-edges.

Conclusion: The field of learning-reasoning approaches in AI research has gained significant attention due to the
advantages it offers by combining neural networks and symbolic reasoning. The integration of neural networks and
symbolic reasoning allows for the utilization of the strengths of both approaches. Neural networks provide the ability
to process complex data and generate predictions, while symbolic reasoning provides a structured and interpretable
framework for representing and reasoning about knowledge. For instance, DeepProbLog and ABL have similar model
principles: the modeling of complex problems is defined in a logic programming language, and the neural network
is used to define simple concepts in a logic programming language. BPGR uses neural networks to accelerate the

Dongran Yu et al.: Preprint submitted to Elsevier Page 18 of 29

 A Survey on Neural-symbolic Learning Systems

search process of symbolic reasoning, along with symbolic knowledge to constrain neural network learning. This
model not only characterizes the matching degree between prediction results and symbolic knowledge but also clearly
states which symbolic knowledge is being fitted, along with the probability of this symbolic knowledge being fitted as
an explanation for the model prediction. However, one limitation of these approaches is their reliance on predefined
logic programming or logic rules, which restricts their generalizability to other tasks. Future research in this field
should explore higher-level interactions between neural networks and symbolic reasoning, such as learning symbolic
knowledge during training. By enhancing the ability of models to acquire and reason with symbolic knowledge in a
data-driven manner, the field can further advance the integration of learning and reasoning in AI systems.
Based on representative works, we summarize a general design idea for neural-symbolic approaches. The
interaction between neural networks and symbolic systems allows encoding the embedding of symbolic knowledge into
neural network models and feeds abstracted symbols of the neural networks into symbolic systems. Further, we provide
some characteristics that should be considered in designing neural-symbolic approaches, as follows: (1) Uncertainty.
The output of the neural network is a distribution, not "True" or "False". Therefore, we need to consider the uncertainty
of the triggered symbolic knowledge. (2) Globalization. It is necessary to consider the fit of all symbolic knowledge
in the knowledge base, not just the local knowledge. (3) Importance. Different knowledge may have different weights,
and the degree of fitting knowledge with different weights should be considered. (4) Interpretability. Interpretability
should be explicitly considered in learning (e.g. the immediate process of the result of learning).

4. Applications
4.1. Object/visual-relationship detection
The goal of object/visual-relationship detection is to recognize objects or the relationships between objects in
images. However, relying solely on visual features to train a model often leads to relatively weak performance. In recent
years, the emergence of neural-symbolic learning systems has paved the way for incorporating external knowledge to
enhance the detection performance of these models. This integration of external knowledge into the learning process
has shown promising results and has become an active area of research in the field.
Donadello et al. [28] propose a novel approach that combines neural networks with first-order logic, known as Logic
Tensor Networks (LTN). By incorporating logical constraints, LTNs enable effective reasoning from noisy images
while also providing a means to describe data characteristics through logic rules. This integration of logic into the neural
network framework enhances interpretability in image recognition tasks. In the context of remote sensing, Marszalek
and Forestier et al. [87, 36] emphasize the utilization of symbolic knowledge from domain experts to improve the
detection capabilities. By leveraging expert knowledge, the remote sensing systems can gain a deeper understanding
of the data and achieve better performance in detecting specific features or patterns. Zhu and Nyga et al. [153, 96]
adopt a different approach by using Markov Logic Networks (MLN) to model symbolic knowledge for integration into
deep learning models. MLNs allow for learning a scoring function and predicting relations between input images and
specific objects or concepts. For example, given an input image of a horse, the model can predict the relation "ridable"
between the horse and people. This approach combines the strengths of deep learning and symbolic reasoning, enabling
more comprehensive and nuanced analysis of visual data.

4.2. Knowledge graph reasoning

Knowledge graphs often suffer from incompleteness, requiring completion or link prediction techniques to enhance
their quality. Zhang et al. [149] provide a comprehensive survey on the benefits of knowledge graph reasoning within
neural-symbolic learning systems. Wang et al. [133] propose a method where triplets or ground rules are transformed
into First-Order Logic (FOL) statements. These FOL statements are then scored using vector/matrix operations based
on the embeddings of the entities and relationships involved. This approach combines logic reasoning with neural
networks to perform link prediction in knowledge graphs. Path-based reasoning approaches [92, 21, 141, 89, 20]
aim to extend reasoning by exploring multi-hop neighbors around a given entity and predicting answers within
these neighborhoods using neural networks. For example, DeepPath [141] employs reinforcement learning to evaluate
sampled paths, reducing the search space and improving efficiency. Building upon path-based reasoning, Teru et al.
[124] propose GraIL, a graph-based reasoning framework. GraIL extracts a subgraph consisting of the k-hop neighbors
of the head and tail entities. Subsequently, a Graph Neural Network (GNN) is employed to reason about the relationship
between the two entities using the extracted subgraph. These approaches demonstrate the integration of neural networks
and graph-based reasoning to perform link prediction and enhance knowledge graph reasoning.

Dongran Yu et al.: Preprint submitted to Elsevier Page 19 of 29

 A Survey on Neural-symbolic Learning Systems

4.3. Classification/ few-shot classification

Marra et al. [83] introduce Relational Neural Machines (RNM), a framework that allows for joint training of learners
and reasoners. This approach enables the integration of both learning and reasoning processes, leading to improved
performance. To address the few-shot learning problem, Sikka et al. [114] incorporate common sense knowledge into
deep neural networks. They also utilize logical knowledge as a neural-symbolic loss function to regularize visual
semantic features. This approach leverages information from unseen classes during model learning, enhancing zero-
shot learning capabilities. Altszyler et al. [3] integrate logic rules into neural network architectures for multi-domain
dialogue recognition tasks. By incorporating logical knowledge, the model can recognize labels of unseen classes
without requiring additional training data. This approach expands the model’s ability to handle previously unseen
classes in the dialogue recognition domain. These methods demonstrate the integration of logical knowledge, common
sense knowledge, and neural networks to improve the performance of various tasks, such as joint training, few-shot
learning, and multi-domain dialogue recognition.

4.4. Intelligent question answering

Indeed, intelligent question answering is a prominent application of neural-symbolic reasoning in natural language
processing and visual reasoning tasks. In this context, the goal is to develop models that can accurately infer answers
to questions by leveraging contextual information from text and images.
Andreas et al. [4] proposed the neural module network (NMN) framework, which uses deep neural network
generating symbolic structures to solve subsequent reasoning problems. Gupta et al. [47] extend NMN and propose an
unsupervised auxiliary loss to help extract arguments associated with the events in the text. Specifically, this method
introduces a reasoning module for the text that enables symbolic reasoning (such as arithmetic, sorting, and counting)
on numbers and dates in a probabilistic or differentiable way, allowing the model to output logical parsing of questions
and intermediate decisions.
Hudson et al. [52] propose a fully differentiable network model (MAC) with cyclic memory, attention, and
composition functions. The MAC provides strong prior conditions for iterative reasoning, transforms the black-box
architecture into something more transparent, and supports interpretability and structured learning. The core concept
is to decompose the image and the question into sequential units, input the recurrent network for sequential reasoning,
and then store the result in the memory unit to calculate the final answer together with the question. Tran and Poon et
al. [129, 101] model domain common sense with MLN and use probabilistic inference methods for the query. Sun et
al. [121] learned a neural semantic parser and trained a model-agnostic model based on meta-learning to improve the
predictive ability of language question-answering tasks cases involving limited simple rules.
Oltramariet et al. [97] proposed integrating neural language models and knowledge graphs in common-sense
question answering. Based on the architecture of the language model, this work proposes an attention-based knowledge
injection method. For visual question-answering tasks, Hudson et al. [53] propose the neural state machine (NSM).
NSM uses a supervised training method to construct a probabilistic scene graph based on the concepts in an image, then
performs sequential reasoning on the probabilistic scene graph, answering questions or discovering new conclusions.

4.5. Reinforcement learning

Deep reinforcement learning is a trending topic in the field of artificial intelligence, and it has been successfully
applied in many contexts. However, current deep reinforcement learning methods have limitations in reasoning ability.
To address this challenge, researchers have started integrating symbolic knowledge into reinforcement learning. In this
paper, we explore two approaches: combining symbolic knowledge with deep reinforcement learning and combining
symbolic knowledge with hierarchical reinforcement learning.
Garnelo et al. [46] proposed a deep symbolic reinforcement learning method (DSRL) that integrates a symbolic
prior into the agent’s learning process to enhance the model’s generalization. The DSRL agent consists of a neural
back end and a symbolic front end. The neural back end learns to map raw sensor data into a symbolic representation,
which is then utilized by the symbolic front end to learn an effective policy. Building upon DSRL, Garcez et al. [39]
extended the approach and introduced a symbolic reinforcement learning method with common sense (SRL+CS). This
method improves both the learning phase and the decision-making phase based on DSRL. In the learning phase, the
reward distribution is no longer based on a fixed calculation formula for updating the Q-values, but instead considers
the interaction between the agent and the object. In the decision-making stage, the model assigns an importance weight
to each Q-function based on the distance between the object and the agent, allowing for a comprehensive aggregation
of the Q-values.

Dongran Yu et al.: Preprint submitted to Elsevier Page 20 of 29

 A Survey on Neural-symbolic Learning Systems

Yang et al. [144] proposed the integration of symbolic planning and hierarchical reinforcement learning (HRL) [7]
to address decision-making in dynamic environments with uncertainties. They introduced a framework called PEORL
(Planning-Execution-Observation-Reinforcement-Learning) that combines these two approaches. Symbolic planning
is employed to guide the agent’s task execution and learning process, while the learned experiences are fed back to
the symbolic knowledge to enhance the planning phase. Specifically, commonsense knowledge of actions constrains
the answer set solver to generate a symbolic plan. The symbolic plan is subsequently mapped to a deterministic
sequence of stochastic options, which guides the hierarchical reinforcement learning (HRL) process. This approach
represents the first utilization of symbolic planning for option discovery within the HRL framework. To achieve
task-level interpretability, Lyu et al. [78] proposed the Symbolic Deep Reinforcement Learning (SDRL) framework,
which shares similarities with REORL and consists of a planner, controller, and meta-controller, along with symbolic
knowledge. The planner employs prior symbolic knowledge to perform long-term planning through a sequence of
symbolic actions (subtasks) that aim to achieve its intrinsic goal. The controller utilizes deep reinforcement learning
(DRL) algorithms to learn sub-policies for each subtask based on intrinsic rewards. The meta-controller learns extrinsic
rewards by evaluating the training performance of the controllers and suggesting new intrinsic goals to the planner.
In essence, both PEORL and SDRL leverage symbolic knowledge to guide the reinforcement learning process and
facilitate decision-making.

5. Future directions
The above paper introduces the current research status and research methods of neural-symbolic learning systems
in detail. On this basis, we discuss some potential future research directions.

5.1. Efficient methods

In neural-symbolic learning systems, symbolic reasoning technologies often encounter challenges related to
intractable precise inference. For instance, in probability inference using Markov Logic Networks (MLNs), the number
of groundings can increase exponentially when dealing with a large number of logic rules and constants. This
exponential growth in grounding can significantly decrease the speed of model inference. While several methods have
been proposed to mitigate this problem, such as learning-based approaches [62, 6], they still have certain limitations.
Approximate inference techniques are commonly employed to improve inference speed but often come at the cost
of reduced accuracy. Consequently, it becomes crucial for researchers to explore the potential of neural networks to
tackle the computational challenges faced by symbolic systems. Designing approaches that leverage the computational
strengths of neural networks to handle tasks that are computationally difficult in traditional symbolic systems represents
a crucial research direction for advancing inference methods.

5.2. Automatic construction of symbolic knowledge

The symbolic knowledge discussed in this paper encompasses logic knowledge and knowledge graphs. Automatic
construction of symbolic knowledge has achieved relative maturity, as evidenced by studies on automatic knowledge
construction [75, 76, 88]. However, the construction of logic rules in neural-symbolic approaches typically relies
on manual efforts from domain experts. This process is time-consuming, laborious, and not easily scalable. A
significant challenge for neural-symbolic learning systems is enabling end-to-end learning for rules that describe prior
knowledge derived from data. While technologies like Inductive Logic Programming (ILP)-based methods exist for
knowledge extraction, the automatic learning of logic rules from data remains largely unexplored. Hence, the automatic
construction of rules represents an important future research direction in the field of neural-symbolic learning systems.

5.3. Symbolic representation learning

A well-designed symbolic representation plays a crucial role in simplifying and improving the efficiency of complex
learning tasks. In the context of zero-shot image classification, for instance, a limited semantic information contained
in a learned symbolic representation can hinder the model’s ability to handle complex classification tasks effectively.
Therefore, precise semantic information encoded in symbolic knowledge is essential for enhancing the performance of
these models. However, most existing symbolic representation learning methods struggle with predicates that exhibit
strong similarity. These predicates may have similar semantics but different logical formulas, such as "next to" and
"near." Current symbolic representation learning methods fail to capture such semantic similarities, which hampers
the reasoning capability of these models. Thus, a significant challenge in the field of neural-symbolic learning systems

Dongran Yu et al.: Preprint submitted to Elsevier Page 21 of 29

 A Survey on Neural-symbolic Learning Systems

is the design of more robust and efficient symbolic representation learning methods. The advancement of graph
representation learning offers a promising avenue for addressing this challenge. By mapping nodes to low-dimensional,
dense, and continuous vectors, graph representation learning can flexibly support various learning and reasoning tasks.
Given that symbolic knowledge often exhibits heterogeneity, multiple relations, and even multimodality, exploring
the development and utilization of heterogeneous graph representation learning methods becomes another important
direction to overcome the challenges faced by neural-symbolic learning systems.

5.4. Application field expansion

Neural-symbolic learning systems have found applications in various domains, including computer vision, natural
language processing, and recommendation systems [154, 104]. Recently, researchers have also started exploring the
potential of applying neural-symbolic learning systems in other areas such as the COVID-19 pandemic [93] and
advanced robotics [16]. For example, in the context of the COVID-19 pandemic, neural-symbolic learning systems
have been utilized for tasks like extracting relevant information from medical literature. Similarly, in the field of
advanced robotics, neural-symbolic learning systems can be employed to enhance robot intelligence and decision-
making capabilities. Given the success of neural-symbolic learning systems in existing domains, it is a natural and
promising direction to extend their application to new domains and develop tailored methods to address specific
challenges in those areas.

6. Conclusion
In this paper, we have presented an overall framework for neural-symbolic learning systems. Our main contribution
is the proposal of a novel taxonomy for neural-symbolic learning systems and the outline of three structured
categorizations. Additionally, we describe the techniques used in each structured categorization, explore a wide range
of applications, and discuss future directions for neural-symbolic learning systems. We firmly believe that a systematic
and comprehensive research survey in this field holds significant value in terms of both theory and application. It
deserves further in-depth research and discussion.

Acknowledgments
This research was partly supported by Foundation item: National Natural Science Foundation of China (61876069);
Jilin Province Key Scientific and Technological Research and Development Project under Grant Nos. 20180201067GX
and 20180201044GX; and Jilin Province Natural Science Foundation (20200201036JC).

CRediT authorship contribution statement

Dongran Yu: Investigation, Data curation, Methodology and Writing. Bo Yang: Conceptualization,Supervision,
Review and Editing. Dayou Liu: Supervision. Hui Wang: Supervision. Shirui Pan: Supervision, Review and Editing.

References
[1] Abboud, R., Ceylan, I., Lukasiewicz, T., 2020. Learning to reason: Leveraging neural networks for approximate dnf counting, in: AAAI, pp.
3097–3104.
[2] Aliseda, A., 2006. Abductive Reasoning: Logical Investigations into Discovery and Explanation.
[3] Altszyler, E., Brusco, P., Basiou, N., Byrnes, J., Vergyri, D., 2021. Zero-shot multi-domain dialog state tracking using descriptive rules, in:
IJCLR.
[4] Andreas, J., Rohrbach, M., Darrell, T., Klein, D., 2016. Neural module networks, in: CVPR, pp. 39–48.
[5] Andrews, R., Diederich, J., Tickle, A.B., 1995. Survey and critique of techniques for extracting rules from trained artificial neural networks,
in: KBS, pp. 373–389.
[6] Bach, S.H., Broecheler, M., Huang, B., Getoor, L., 2015. Hinge-loss markov random fields and probabilistic soft logic, in: arXiv preprint
arXiv:1505.04406.
[7] Barto, A.G., Mahadevan, S., 2003. Recent advances in hierarchical reinforcement learning, in: DEDS, pp. 41–77.
[8] Besold, T.R., Garcez, A.d., Bader, S., Bowman, H., Domingos, P., Hitzler, P., Kühnberger, K.U., Lamb, L.C., Lowd, D., Lima, P.M.V., et al.,
2017. Neural-symbolic learning and reasoning: A survey and interpretation, in: arXiv preprint arXiv:1711.03902.
[9] Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J., Yakhnenko, O., 2013. Translating embeddings for modeling multi-relational data, in:
NIPS.
[10] Bosselut, A., Rashkin, H., Sap, M., Malaviya, C., Celikyilmaz, A., Choi, Y., 2019. Comet: Commonsense transformers for automatic
knowledge graph construction, in: ACL.

Dongran Yu et al.: Preprint submitted to Elsevier Page 22 of 29

 A Survey on Neural-symbolic Learning Systems

[11] Cai, L.W., Dai, W.Z., Huang, Y.X., Li, Y.F., Muggleton, S., Jiang, Y., 2021. Abductive learning with ground knowledge base, in: IJCAI.
[12] Calegari, R., Ciatto, G., Omicini, A., 2020. On the integration of symbolic and sub-symbolic techniques for xai: A survey, in: IA, pp. 7–32.
[13] Campero, A., Pareja, A., Klinger, T., Tenenbaum, J., Riedel, S., 2018. Logical rule induction and theory learning using neural theorem
proving, in: NIPS.
[14] Chen, R., Chen, T., Hui, X., Wu, H., Li, G., Lin, L., 2020. Knowledge graph transfer network for few-shot recognition, in: AAAI, pp.
10575–10582.
[15] Chen, T., Chen, R., Nie, L., Luo, X., Liu, X., Lin, L., 2018. Neural task planning with and–or graph representations, in: IEEE Transactions
on Multimedia, pp. 1022–1034.
[16] Conti, C.J., Varde, A.S., Wang, W., 2020. Robot action planning by commonsense knowledge in human-robot collaborative tasks, in:
IEMTRONICS, pp. 1–7.
[17] Darwiche, A., 2001. On the tractable counting of theory models and its application to truth maintenance and belief revision, in: JANCL, pp.
11–34.
[18] Darwiche, A., 2011. Sdd: A new canonical representation of propositional knowledge bases, in: AI.
[19] Darwiche, A., Marquis, P., 2002. A knowledge compilation map, in: JAIR, pp. 229–264.
[20] Das, R., Dhuliawala, S., Zaheer, M., Vilnis, L., Durugkar, I., Krishnaurthy, A., Smola, A., McCallum, A., 2017. Go for a walk and arrive at
the answer: Reasoning over knowledge bases with reinforcement learning., in: NIPS.
[21] Das, R., Neelakantan, A., Belanger, D., McCallum, A., 2016. Chains of reasoning over entities, relations, and text using recurrent neural
networks, in: ACL.
[22] Davis, E., 2017. Logical formalizations of commonsense reasoning: a survey, in: JAIR, pp. 651–723.
[23] De Raedt, L., Kimmig, A., Toivonen, H., 2007. Problog: A probabilistic prolog and its application in link discovery., in: IJCAI, pp. 2462–2467.
[24] Dettmers, T., Minervini, P., Stenetorp, P., Riedel, S., 2018. Convolutional 2d knowledge graph embeddings, in: AAAI.
[25] Diligenti, M., Gori, M., Sacca, C., 2017. Semantic-based regularization for learning and inference, in: AI, pp. 143–165.
[26] Dollár, K.H.G.G.P., Girshick, R., 2017. Mask r-cnn, in: ICCV, pp. 2961–2969.
[27] Domingos, P., Lowd, D., 2019. Unifying logical and statistical ai with markov logic, in: Commun.ACM, pp. 74–83.
[28] Donadello, I., Serafini, L., Garcez, A.D., 2017. Logic tensor networks for semantic image interpretation, in: IJCAI.
[29] Dong, H., Mao, J., Lin, T., Wang, C., Li, L., Zhou, D., 2019. Neural logic machines, in: ICLR.
[30] Dos Martires, P.Z., Derkinderen, V., Manhaeve, R., Meert, W., Kimmig, A., De Raedt, L., 2019. Transforming probabilistic programs into
algebraic circuits for inference and learning, in: NIPS.
[31] Dragone, P., Teso, S., Passerini, A., 2021. Neuro-symbolic constraint programming for structured prediction, in: arXiv preprint
arXiv:2103.17232.
[32] Eisner, J., 2002. Parameter estimation for probabilistic finite-state transducers, in: ACL, pp. 1–8.
[33] Ellis, K.M., Morales, L.E., Sablé-Meyer, M., Solar Lezama, A., Tenenbaum, J.B., 2018. Library learning for neurally-guided bayesian
program induction, in: NIPS.
[34] Enderton, H.B., 2001. A mathematical introduction to logic.
[35] Evans, R., Grefenstette, E., 2018. Learning explanatory rules from noisy data, in: JAIR, pp. 1–64.
[36] Forestier, G., Wemmert, C., Puissant, A., 2013. Coastal image interpretation using background knowledge and semantics, in: Comput Geosci,
pp. 88–96.
[37] Galárraga, L., Teflioudi, C., Hose, K., Suchanek, F.M., 2015. Fast rule mining in ontological knowledge bases with amie + +, in: VLDB,
pp. 707–730.
[38] Garcez, A.d., Besold, T.R., De Raedt, L., Földiak, P., Hitzler, P., Icard, T., Kühnberger, K.U., Lamb, L.C., Miikkulainen, R., Silver, D.L.,
2015. Neural-symbolic learning and reasoning: contributions and challenges, in: AAAI.
[39] Garcez, A.d., Dutra, A.R.R., Alonso, E., 2018. Towards symbolic reinforcement learning with common sense, in: arXiv preprint
arXiv:1804.08597.
[40] Garcez, A.d., Gori, M., Lamb, L.C., Serafini, L., Spranger, M., Tran, S.N., 2019. Neural-symbolic computing: An effective methodology for
principled integration of machine learning and reasoning, in: arXiv preprint arXiv:1905.06088.
[41] Garcez, A.d., Lamb, L.C., 2020. Neurosymbolic ai: the 3rd wave, in: arXiv preprint arXiv:2012.05876.
[42] Garcez, A.S.A., Zaverucha, G., 1999. The connectionist inductive learning and logic programming system, in: APIN, pp. 59–77.
[43] Garcez, A.S.d., Broda, K., Gabbay, D.M., et al., 2002. Neural-symbolic learning systems: foundations and applications.
[44] Garcez, A.S.d., Broda, K.B., Gabbay, D.M., 2012. Neural-symbolic learning systems: foundations and applications.
[45] Gardner, M., Mitchell, T., 2015. Efficient and expressive knowledge base completion using subgraph feature extraction, in: EMNLP, pp.
1488–1498.
[46] Garnelo, M., Arulkumaran, K., Shanahan, M., 2016. Towards deep symbolic reinforcement learning, in: NIPS.
[47] Gupta, N., Lin, K., Roth, D., Singh, S., Gardner, M., 2020. Neural module networks for reasoning over text, in: ICLR.
[48] Gupta, V., Patro, B.N., Parihar, H., Namboodiri, V.P., 2022. Vquad: Video question answering diagnostic dataset, in: WACVW, pp. 282–291.
[49] Hoffmann, J., Navarro, O., Kastner, F., Janßen, B., Hubner, M., 2017. A survey on cnn and rnn implementations, in: PESARO.
[50] Honavar, V., 1995. Symbolic artificial intelligence and numeric artificial neural networks: towards a resolution of the dichotomy.
Computational architectures integrating neural and symbolic processes: a perspective on the state of the art , 351–388.
[51] Hu, Z., Ma, X., Liu, Z., Hovy, E., Xing, E., 2016. Harnessing deep neural networks with logic rules, in: ACL.
[52] Hudson, D.A., Manning, C.D., 2018. Compositional attention networks for machine reasoning, in: ICLR.
[53] Hudson, D.A., Manning, C.D., 2019. Learning by abstraction: The neural state machine, in: NIPS.
[54] Ji, J., Zhu, F., Cui, J., Zhao, H., Yang, B., 2022. A dual-system method for intelligent fault localization in communication networks, in: ICC
2022-IEEE International Conference on Communications, pp. 4062–4067.
[55] Jiang, X., Wang, Q., Wang, B., 2019. Adaptive convolution for multi-relational learning, in: NAACL HLT, pp. 978–987.

Dongran Yu et al.: Preprint submitted to Elsevier Page 23 of 29

 A Survey on Neural-symbolic Learning Systems

[56] Kahneman, D., 2011. Thinking, fast and slow.

[57] Kalyan, A., Mohta, A., Polozov, O., Batra, D., Jain, P., Gulwani, S., 2018. Neural-guided deductive search for real-time program synthesis
from examples, in: ICLR.
[58] Kampffmeyer, M., Chen, Y., Liang, X., Wang, H., Zhang, Y., Xing, E.P., 2019. Rethinking knowledge graph propagation for zero-shot
learning, in: CVPR, pp. 11487–11496.
[59] Karpas, E., Abend, O., Belinkov, Y., Lenz, B., Lieber, O., Ratner, N., Shoham, Y., Bata, H., Levine, Y., Leyton-Brown, K., et al., 2022. Mrkl
systems: A modular, neuro-symbolic architecture that combines large language models, external knowledge sources and discrete reasoning,
in: arXiv preprint arXiv:2205.00445.
[60] Kaur, N., Kunapuli, G., Khot, T., Kersting, K., Cohen, W., Natarajan, S., 2017. Relational restricted boltzmann machines: A probabilistic
logic learning approach, in: IJCLR, pp. 94–111.
[61] Kautz, H., 2022. The third ai summer: Aaai robert s. engelmore memorial lecture. AI Magazine 43, 105–125.
[62] Khot, T., Natarajan, S., Kersting, K., Shavlik, J., 2011. Learning markov logic networks via functional gradient boosting, in: ICDMW, pp.
320–329.
[63] Kipf, T.N., Welling, M., 2017. Semi-supervised classification with graph convolutional networks, in: ICLR.
[64] Koller, D., Friedman, N., Džeroski, S., Sutton, C., McCallum, A., Pfeffer, A., Abbeel, P., Wong, M.F., Meek, C., Neville, J., et al., 2007.
Introduction to statistical relational learning.
[65] Lamb, L.C., Garcez, A., Gori, M., Prates, M., Avelar, P., Vardi, M., 2020. Graph neural networks meet neural-symbolic computing: A survey
and perspective, in: AAAI.
[66] Landajuela, M., Petersen, B.K., Kim, S., Santiago, C.P., Glatt, R., Mundhenk, N., Pettit, J.F., Faissol, D., 2021. Discovering symbolic policies
with deep reinforcement learning, in: ICML, pp. 5979–5989.
[67] Lavrac, N., Dzeroski, S., 1994. Inductive logic programming., in: WLP, pp. 146–160.
[68] LeCun, Y., Bottou, L., Bengio, Y., Haffner, P., 1998. Gradient-based learning applied to document recognition, in: Proceedings of the IEEE,
pp. 2278–2324.
[69] Levine, Y., Dalmedigos, I., Ram, O., Zeldes, Y., Jannai, D., Muhlgay, D., Osin, Y., Lieber, O., Lenz, B., Shalev-Shwartz, S., et al., 2022.
Standing on the shoulders of giant frozen language models, in: arXiv preprint arXiv:2204.10019.
[70] Li, A., Luo, T., Lu, Z., Xiang, T., Wang, L., 2019. Large-scale few-shot learning: Knowledge transfer with class hierarchy, in: CVPR, pp.
7212–7220.
[71] Li, Y., Tarlow, D., Brockschmidt, M., Zemel, R., 2016. Gated graph sequence neural networks, in: ICLR.
[72] Liang, C., Berant, J., Le, Q., Forbus, K.D., Lao, N., 2017. Neural symbolic machines: Learning semantic parsers on freebase with weak
supervision, in: ACL.
[73] Lin, Y., Liu, Z., Sun, M., Liu, Y., Zhu, X., 2015. Learning entity and relation embeddings for knowledge graph completion, in: AAAI.
[74] Liu, Z., Jiang, Z., Wei, F., 2019. Od-gcn object detection by knowledge graph with gcn, in: arXiv preprint arXiv:1908.04385.
[75] LiuQiao, L., DuanHong, L., et al., 2016. Knowledge graph construction techniques, in: computer research and development, p. 582.
[76] Luan, Y., He, L., Ostendorf, M., Hajishirzi, H., 2018. Multi-task identification of entities, relations, and coreference for scientific knowledge
graph construction, in: EMNLP.
[77] Luo, R., Zhang, N., Han, B., Yang, L., 2020. Context-aware zero-shot recognition, in: AAAI, pp. 11709–11716.
[78] Lyu, D., Yang, F., Liu, B., Gustafson, S., 2019. Sdrl: interpretable and data-efficient deep reinforcement learning leveraging symbolic
planning, in: AAAI, pp. 2970–2977.
[79] Manhaeve, R., De Raedt, L., Kimmig, A., Dumancic, S., Demeester, T., 2019. Deepproblog: Integrating logic and learning through algebraic
model counting, in: NIPS.
[80] Manhaeve, R., Dumančić, S., Kimmig, A., Demeester, T., De Raedt, L., 2018. Deepproblog: Neural probabilistic logic programming, in:
NeurIPS.
[81] Mao, J., Gan, C., Kohli, P., Tenenbaum, J.B., Wu, J., 2019. The neuro-symbolic concept learner: Interpreting scenes, words, and sentences
from natural supervision, in: arXiv preprint arXiv:1904.12584.
[82] Marcus, G., 2020. The next decade in ai: four steps towards robust artificial intelligence. arXiv preprint arXiv:2002.06177 .
[83] Marra, G., Diligenti, M., Giannini, F., Gori, M., Maggini, M., 2020a. Relational neural machines.
[84] Marra, G., Dumančić, S., Manhaeve, R., De Raedt, L., 2020b. From statistical relational to neural symbolic artificial intelligence: a survey,
in: IJCAI.
[85] Marra, G., Giannini, F., Diligenti, M., Gori, M., 2019. Integrating learning and reasoning with deep logic models, in: ECML PKDD.
[86] Marra, G., Kuželka, O., 2021. Neural markov logic networks, in: UAI.
[87] Marszalek, M., Schmid, C., 2007. Semantic hierarchies for visual object recognition, in: CVPR, pp. 1–7.
[88] Martinez-Rodriguez, J.L., López-Arévalo, I., Rios-Alvarado, A.B., 2018. Openie-based approach for knowledge graph construction from
text, in: Expert Syst. Appl., pp. 339–355.
[89] Meilicke, C., Chekol, M.W., Fink, M., Stuckenschmidt, H., 2020. Reinforced anytime bottom up rule learning for knowledge graph
completion, in: arXiv preprint arXiv:2004.04412.
[90] Mihalkova, L., Mooney, R.J., 2007. Bottom-up learning of markov logic network structure, in: ICML, pp. 625–632.
[91] Minervini, P., Demeester, T., Rocktäschel, T., Riedel, S., 2017. Adversarial sets for regularising neural link predictors, in: arXiv preprint
arXiv:1707.07596.
[92] Neelakantan, A., Roth, B., McCallum, A., 2015. Compositional vector space models for knowledge base inference, in: AAAI.
[93] Ngan, K.H., Garcez, A.D., Townsend, J., 2022. Extracting meaningful high-fidelity knowledge from convolutional neural networks, in:
IJCNN, pp. 1–17.
[94] Novák, V., Perfilieva, I., Mockor, J., 2012. Mathematical principles of fuzzy logic. volume 517.
[95] Nye, M., Hewitt, L., Tenenbaum, J., Solar-Lezama, A., 2019. Learning to infer program sketches, in: ICML, pp. 4861–4870.

Dongran Yu et al.: Preprint submitted to Elsevier Page 24 of 29

 A Survey on Neural-symbolic Learning Systems

[96] Nyga, D., Balint-Benczedi, F., Beetz, M., 2014. Pr2 looking at things—ensemble learning for unstructured information processing with
markov logic networks, in: ICRA, pp. 3916–3923.
[97] Oltramari, A., Francis, J., Ilievski, F., Ma, K., Mirzaee, R., 2021. Generalizable neuro-symbolic systems for commonsense question
answering, in: Neuro-Symbolic Artificial Intelligence: The State of the Art, pp. 294–310.
[98] Payani, A., Fekri, F., 2019. Inductive logic programming via differentiable deep neural logic networks, in: arXiv preprint arXiv:1906.03523.
[99] Perotti, A., Boella, G., Colombo Tosatto, S., d’Avila Garcez, A.S., Genovese, V., van der Torre, L., 2012. Learning and reasoning about
norms using neural-symbolic systems, in: AAMAS, pp. 1023–1030.
[100] Poon, H., Domingos, P., 2006. Sound and efficient inference with probabilistic and deterministic dependencies, in: AAAI, pp. 458–463.
[101] Poon, H., Domingos, P., 2009. Unsupervised semantic parsing, in: EMNLP, pp. 1–10.
[102] Prates, M., Avelar, P.H., Lemos, H., Lamb, L.C., Vardi, M.Y., 2019. Learning to solve np-complete problems: A graph neural network for
decision tsp, in: AAAI, pp. 4731–4738.
[103] Qu, M., Tang, J., 2020. Probabilistic logic neural networks for reasoning, in: ICLR.
[104] Raizada, M., 2022. Survey on recommender systems incorporating trust, in: ICAAIC, pp. 1011–1015.
[105] Ratti, E., Graves, M., 2022. Explainable machine learning practices: opening another black box for reliable medical ai, in: AI and Ethics, pp.
1–14.
[106] Richardson, M., Domingos, P., 2006. Markov logic networks, in: ML, pp. 107–136.
[107] Riegel, R., Gray, A., Luus, F., Khan, N., Makondo, N., Akhalwaya, I.Y., Qian, H., Fagin, R., Barahona, F., Sharma, U., et al., 2020. Logical
neural networks, in: arXiv preprint arXiv:2006.13155.
[108] Rissati, J.V., Molina, P.C., Anjos, C.S., 2020. Hyperspectral image classification using random forest and deep learning algorithms, in:
LAGIRS, pp. 132–132.
[109] Rocktäschel, T., Riedel, S., 2017. End-to-end differentiable proving, in: NIPS.
[110] Safavian, S.R., Landgrebe, D., 1991. A survey of decision tree classifier methodology, in: IEEE T SYST MAN CY-S, pp. 660–674.
[111] Salahuddin, Z., Woodruff, H.C., Chatterjee, A., Lambin, P., 2022. Transparency of deep neural networks for medical image analysis: A
review of interpretability methods, in: CIBM, pp. 105–111.
[112] Schlichtkrull, M., Kipf, T.N., Bloem, P., Berg, R.v.d., Titov, I., Welling, M., 2018. Modeling relational data with graph convolutional networks,
in: ESWC, pp. 593–607.
[113] Serafini, L., Garcez, A.d., 2016. Logic tensor networks: Deep learning and logical reasoning from data and knowledge, in: arXiv preprint
arXiv:1606.04422.
[114] Sikka, K., Huang, J., Silberfarb, A., Nayak, P., Rohrer, L., Sahu, P., Byrnes, J., Divakaran, A., Rohwer, R., 2020. Zero-shot learning with
knowledge enhanced visual semantic embeddings, in: arXiv preprint arXiv:2011.10889.
[115] Silva, A., Gombolay, M., 2021. Encoding human domain knowledge to warm start reinforcement learning, in: AAAI, pp. 5042–5050.
[116] Singla, P., Domingos, P., 2005. Discriminative training of markov logic networks, in: AAAI, pp. 868–873.
[117] Singla, P., Domingos, P., 2006. Memory-efficient inference in relational domains, in: AAAI, pp. 488–493.
[118] Sourek, G., Aschenbrenner, V., Zelezny, F., Schockaert, S., Kuzelka, O., 2018. Lifted relational neural networks: Efficient learning of latent
relational structures, in: JAIR, pp. 69–100.
[119] Sun, R., Alexandre, F., 2013. Connectionist-symbolic integration: From unified to hybrid approaches.
[120] Sun, R., Bookman, L.A., 1994. Computational architectures integrating neural and symbolic processes: A perspective on the state of the art .
[121] Sun, Y., Tang, D., Duan, N., Gong, Y., Feng, X., Qin, B., Jiang, D., 2020. Neural semantic parsing in low-resource settings with back-
translation and meta-learning, in: AAAI, pp. 8960–8967.
[122] Sun, Z., Deng, Z.H., Nie, J.Y., Tang, J., 2019. Rotate: Knowledge graph embedding by relational rotation in complex space, in: ICLR.
[123] Tandon, N., Varde, A.S., de Melo, G., 2018. Commonsense knowledge in machine intelligence, in: SIGMOD, pp. 49–52.
[124] Teru, K., Denis, E., Hamilton, W., 2020. Inductive relation prediction by subgraph reasoning, in: ICML, pp. 9448–9457.
[125] Tian, J., Li, Y., Chen, W., Xiao, L., He, H., Jin, Y., 2022. Weakly supervised neural symbolic learning for cognitive tasks, in: AAAI.
[126] Towell, G.G., Shavlik, J.W., 1994. Knowledge-based artificial neural networks, in: AI, pp. 119–165.
[127] Townsend, J., Chaton, T., Monteiro, J.M., 2019. Extracting relational explanations from deep neural networks: A survey from a neural-
symbolic perspective, in: TNNLS, pp. 3456–3470.
[128] Townsend, J., Chaton, T., Monteiro, J.M., 2020. Extracting relational explanations from deep neural networks: A survey from a neural-
symbolic perspective, in: TNNLS, pp. 3456–3470.
[129] Tran, S.D., Davis, L.S., 2008. Event modeling and recognition using markov logic networks, in: ECCV, pp. 610–623.
[130] Trouillon, T., Welbl, J., Riedel, S., Gaussier, É., Bouchard, G., 2016. Complex embeddings for simple link prediction, in: ICML, pp. 2071–
2080.
[131] Vashishth, S., Sanyal, S., Nitin, V., Talukdar, P., 2020. Composition-based multi-relational graph convolutional networks, in: ICLR.
[132] Von Rueden, L., Mayer, S., Beckh, K., Georgiev, B., Giesselbach, S., Heese, R., Kirsch, B., Pfrommer, J., Pick, A., Ramamurthy, R., et al.,
2021. Informed machine learning–a taxonomy and survey of integrating prior knowledge into learning systems, in: TKDE, pp. 614–633.
[133] Wang, P., Dou, D., Wu, F., de Silva, N., Jin, L., 2019a. Logic rules powered knowledge graph embedding, in: arXiv preprint arXiv:1903.03772.
[134] Wang, X., Ye, Y., Gupta, A., 2018. Zero-shot recognition via semantic embeddings and knowledge graphs, in: CVPR, pp. 6857–6866.
[135] Wang, Y., Yao, Q., Kwok, J.T., Ni, L.M., 2020. Generalizing from a few examples: A survey on few-shot learning, in: CSUR, pp. 1–34.
[136] Wang, Z., Ren, Z., He, C., Zhang, P., Hu, Y., 2019b. Robust embedding with multi-level structures for link prediction., in: IJCAI, pp.
5240–5246.
[137] Wang, Z., Zhang, J., Feng, J., Chen, Z., 2014. Knowledge graph embedding by translating on hyperplanes, in: AAAI.
[138] Wen, L.H., Jo, K.H., 2022. Deep learning-based perception systems for autonomous driving: A comprehensive survey, in: Neurocomputing.
[139] Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., Philip, S.Y., 2020. A comprehensive survey on graph neural networks, in: TNNLS, pp. 4–24.
[140] Xie, Y., Xu, Z., Kankanhalli, M.S., Meel, K.S., Soh, H., 2019. Embedding symbolic knowledge into deep networks .

Dongran Yu et al.: Preprint submitted to Elsevier Page 25 of 29

 A Survey on Neural-symbolic Learning Systems

[141] Xiong, W., Hoang, T., Wang, W.Y., 2017. Deeppath: A reinforcement learning method for knowledge graph reasoning, in: EMNLP.
[142] Xu, J., Zhang, Z., Friedman, T., Liang, Y., Broeck, G., 2018. A semantic loss function for deep learning with symbolic knowledge, in: ICML,
pp. 5502–5511.
[143] Yang, B., Yih, W.t., He, X., Gao, J., Deng, L., 2015. Embedding entities and relations for learning and inference in knowledge bases, in:
ICLR.
[144] Yang, F., Lyu, D., Liu, B., Gustafson, S., 2018. Peorl: Integrating symbolic planning and hierarchical reinforcement learning for robust
decision-making, in: IJCAI.
[145] Yang, F., Yang, Z., Cohen, W.W., 2017. Differentiable learning of logical rules for knowledge base completion, in: NIPS.
[146] Yang, Y., Song, L., 2020. Learn to explain efficiently via neural logic inductive learning, in: ICLR.
[147] Yi, K., Wu, J., Gan, C., Torralba, A., Kohli, P., Tenenbaum, J.B., 2018. Neural-symbolic vqa: Disentangling reasoning from vision and
language understanding, in: NIPS.
[148] Yu, D., Yang, B., Wei, Q., Li, A., Pan, S., 2022. A probabilistic graphical model based on neural-symbolic reasoning for visual relationship
detection, in: CVPR, pp. 10609–10618.
[149] Zhang, J., Chen, B., Zhang, L., Ke, X., Ding, H., 2021. Neural, symbolic and neural-symbolic reasoning on knowledge graphs, in: AI Open,
pp. 14–35.
[150] Zhang, Y., Chen, X., Yang, Y., Ramamurthy, A., Li, B., Qi, Y., Song, L., 2020. Efficient probabilistic logic reasoning with graph neural
networks, in: ICLR.
[151] Zhong, J., Haoran, W., Yunlong, Y., Yanwei, P., 2019. A decadal survey of zero-shot image classification, in: SCIENTIA SINICA
Informationis, pp. 1299–1320.
[152] Zhou, Z.H., 2019. Abductive learning: towards bridging machine learning and logical reasoning, in: Science China Information Sciences,
pp. 1–3.
[153] Zhu, Y., Fathi, A., Fei-Fei, L., 2014. Reasoning about object affordances in a knowledge base representation, in: ECCV, pp. 408–424.
[154] Zhu, Y., Xian, Y., Fu, Z., de Melo, G., Zhang, Y., 2021. Faithfully explainable recommendation via neural logic reasoning, in: ACL.
[155] Zuidberg Dos Martires, P., Kumar, N., Persson, A., Loutfi, A., De Raedt, L., 2020. Symbolic learning and reasoning with noisy data for
probabilistic anchoring, in: Front. Robot. AI, p. 100.

A. Preliminaries
In this section, we introduce background information related to symbolic knowledge and neural networks. For
symbolic knowledge, we focus on two categories: logic knowledge and knowledge graphs. Logic knowledge can be
further subdivided into propositional logic and first-order logic.

A.1. Symbols
A.1.1. Propositional logic
Propositional logic statements are declarative sentences that are either True or False. A declarative sentence is a
True sentence if it is consistent with the fact; otherwise, it is a False sentence. The connectors between propositions
are "∧", "∨", "¬" and "⇒". Propositional logic can be expressed in the form of the following formula:

𝑃 ⇒ 𝑄, (9)

where 𝑃 represents the antecedent (condition), while 𝑄 represents the consequent (conclusion).
Propositional logic is usually compiled in the form of directed acyclic graphs. Conjunctive Normal Forms (CNFs),
deterministic-Decomposable Negation Normal Forms (d-DNNFs) [17, 19], and Sentential Decision Diagrams (SDDs)
[18] are representative examples of knowledge representation, where SDD is a subset of d-DNNFs. For example, given
a propositional logic 𝑆𝑚𝑜𝑘𝑒𝑠 ⇒ 𝐶𝑜𝑢𝑔ℎ, its CNF and d-DNNF graph are presented in Figure 16 (a) and Figure 16 (b),
respectively.

Table 4
An instance of a Markov logical network.
Proposition First-order logic Weight
Smoking causes cough. 𝐹 1 ∶ ∀𝑥, 𝑆𝑚𝑜𝑘𝑒𝑠(𝑥) ⇒ 𝐶𝑜𝑢𝑔ℎ(𝑥) 1.5
If two people are friends, either both smoke or neither does. 𝐹 2 ∶ ∀𝑥∀𝑦, 𝐹 𝑟𝑖𝑒𝑛𝑑𝑠(𝑥, 𝑦) ⇒ (𝑆𝑚𝑜𝑘𝑒𝑠(𝑥) ⇔ 𝑆𝑚𝑜𝑘𝑒𝑠(𝑦)) 1.1

A.1.2. First-order logic

Propositional logic is hard to describe complex problems; for these, predicate logic is required. In this paper, we
introduce one of the predicate logic –first-order logic (FOL) [34]. FOL consists of four types of elements, connectors,

Dongran Yu et al.: Preprint submitted to Elsevier Page 26 of 29

 A Survey on Neural-symbolic Learning Systems

^ ^
⋁
⋁

^
-Smokes Cough -Smokes Cough Smokes

（a）CNF （b）d-DNNF
Figure 16: Directed acyclic graph of CNF and d-DNNF. In the graph, leaf nodes represent atoms of propositional logic,
and other nodes represent connectors. Directed edges are the relationship between nodes.

and quantifiers. The four types include constants, variables, functions, and predicates. Constants represent objects in
the domain of interest (for example, in the predicate father(a,b), a=Bob, b=Mara, a and b are constant). Variables
range over the objects in the domain (for example, in the predicate father(𝑥,𝑦), where 𝑥 is the father of 𝑦, and the
variable 𝑥 is limited to the scope of the father class). Functions represent mappings from tuples of objects to objects.
Predicates represent relations among objects in a given domain or attributes of these objects. The connector is the same
as in propositional logic. FOL involves the combination of atoms through connectors, such that an expression can be
written in the following form:

𝐵1 (𝑥) ∧ 𝐵2 (𝑥) ∧ ⋯ ∧ 𝐵𝑛 (𝑥) ⇒ 𝐻(𝑥), (10)

where 𝐵1 (𝑥), 𝐵2 (𝑥), ⋯, 𝐵𝑛 (𝑥) represents the rule body, which is composed of multiple atoms. 𝐻(𝑥) represents the
rule head and is the result derived from the rule body.
Knowledge representation of FOL can be achieved by a Markov logic network (MLN) [106]. MLN is an undirected
graph in which each node represents a variable, and the joint distribution is represented as follows:
∑
𝑃 (𝑋 = 𝑥) = 1∕𝑍𝑒𝑥𝑝{ 𝑤𝑖 𝑛𝑖 (𝑥)}, (11)
𝑖

where 𝑍 represents the partition function, 𝑤𝑖 represents the weight of the rule, 𝑛𝑖 (𝑥) represents the number of times
that the value of the rule is true, and we use t-norm fuzzy logic [94] to calculate logical connectives.
The following introduces a simple example of an MLN. Table 4 shows the two rules (𝐹1 ,1.5), (𝐹2 ,1.1) of this
example [27]. Given a constant set 𝐶 = {𝐴, 𝐵}, the generated ground Markov logic network is shown in Figure 17.

A.1.3. Knowledge graph

A knowledge graph is a directed and labeled graph. Nodes in this graph represent semantic symbols (concepts),
such as animals, computers, people, etc; for their part, edges connect node pairs and express the semantic relationships
between them, such as the food chain relationship between animals, friend relationship between people, etc. Knowledge
graphs can be formally expressed in the form: 𝐺 = (𝐻, 𝑅, 𝑇 ), here, 𝐻 = {ℎ1 , ℎ2 , ..., ℎ𝑛 } represents the set of head
entities, and 𝑛 represents the number of head entities, 𝑇 = {𝑡1 , 𝑡2 , ..., 𝑡𝑚 } represents the set of tail entities, and 𝑚
represents the number of tail entities; moreover, 𝑅 = {𝑟1 , 𝑟2 , ..., 𝑟𝑘 } represents the set of relationships, and 𝑘 represents
the number of relationships. Figure 18 presents an example. For a given triplet (cat, attribute, paw), the nodes (head
entity and tail entity) are cat and paw, and the relationship is an attribute.

Dongran Yu et al.: Preprint submitted to Elsevier Page 27 of 29

 A Survey on Neural-symbolic Learning Systems

Friends(A,B)

Friends(A,A) Smokes(A) Smokes(B) Friends(B,B)

Cough(A)
Cough(B)

Friends(B,A)

Figure 17: Ground Markov logic network. Nodes are variables and edges are the relationship between variables.

Panda
Bird

Tail Mammals

Cat Horse
Ears
Attribute

Paw
Dog
Knowledge graph

Figure 18: An example of a knowledge graph.

For their part, a knowledge graph representation is used to encode discrete symbols (entities, attributes, relation-
ships, etc.) into a low-dimensional vector space to obtain a distributed representation. Typical methods include R-
GCN [112], M-GNN [136], CompGCN [131], TransE [9],TransR [73], TransH [137], RotatE [122], DisMult [143],
ComplEX [130], ConvE [24], ConvR [55], GGNN [71], and GCN [63], among others.

A.2. Neural networks

Although neural networks exist in many forms, we only introduce those relevant to works discussed in Section 3.
CNNs (Convolution Neural Networks, Figure 19) [68] have a set of convolutional layers that extracts features
from the input (The input usually is the image). Next, one or more fully-connected layers are applied to extract the
classification or regression information with the help of logistic or linear regression. Usually, the final layer of a CNN
is a "softmax" layer, where the sum of all activations is one.

Dongran Yu et al.: Preprint submitted to Elsevier Page 28 of 29

 A Survey on Neural-symbolic Learning Systems
Class 1
Class 1
Node Feature map Class 2
channel embedding Class 2

…
Image softmax

…
Input Class n
Class n
Input layers Hidden layers Output layers
Convolutional layers Fully-connected layers

Figure 19: Diagram of a CNN.

Graph structure
Task-related outputs
Node feature

Input Graph convolutional layers Task-related modules

Figure 20: GNN, which feeds graph structure and node feature into graph convolutional layers.

GNNs (Graph Neural Networks, Figure 20) [139] with multiple graph convolutional layers that can attain each
node’s hidden representation by aggregating feature information from its neighbors. The graph structure and node
feature information as inputs and the outputs of GNNs can focus on different graph analysis tasks, such as node
classification and graph classification. Therefore, GNNs have different variants, such as GCN, GAE, etc.

Input 1 Output 1

Input 2 Output 2
…

…
…

Input n Output n

Hidden layers

Figure 21: RNN, which feeds its output back into its input. Here, dashed lines represent recurrent connections.

RNNs (Recursive Neural Networks, Figure 21) [49] are neural networks in which the input and output layers
are identical, and the latter propagates activation back to the former through recurrent connections. Over time, with
each time step consisting of a full input-to-output forward pass, the hidden layer transitions through a series of states.
Usually, the input is sequence data such as sentences in natural language processing. LSTM (Long Short-Term Memory
Networks) is a variable of RNN.

Dongran Yu et al.: Preprint submitted to Elsevier Page 29 of 29