1

Understanding AI, Part 1: Data Training

Sylvia He
(Written in 2021 for Mouser)

What is AI and Machine Learning?

Artificial intelligence (AI), or machine intelligence, informally refers to a machine’s ability to mimic
human cognitive functions, such as learning, perceiving, reasoning, and problem-solving.
With the human-like abilities to “think”, AI sees action in diverse tasks in many sectors, such as image
recognition in medical diagnosis and security surveillance, speech recognition, translation, drug
discovery and development, and financial fraud detection. With the exponential growth of the Internet
of Things (IoT), AI will be crucial in operating IoT devices like autonomous vehicles, surgical robots,
and military drones.

Inspired by the brain, deep learning is a type of machine learning that converts images, voices, or
texts into numbers and then analyzes with multiple layers of mathematical manipulations (hence the
description “deep learning”). In the process, the layers of analysis form a hierarchy of concepts or a
system of “thinking”, that can deconstruct an input and identify the underlying trends or patterns.
Deep learning also diverges from the brain in many ways. For example, the brain has different types
of neurons and functional zones, while machine learning is currently focused on categorizing
information with the highest level of precision.
Like humans, AI needs to learn a task before doing it. As the human brain learns from external cues
to establish systems of thinking to solve future problems, a computer algorithm (a problem-solving
process or a set of rules) learns from data and generates models (or a set of parameters like weights
and biases) that make future problem-solving as efficient and accurate as possible (or “minimum
error, or minimum loss”).
The many types of machine learning fall into the categories of supervised, transfer, unsupervised,
semi-supervised, ensemble, and reinforcement.

Supervised Learning

If a student is learning under supervision, the teacher needs to ensure the student is learning correctly
and achieving high accuracy. Similarly, in supervised learning, an algorithm learns with a complete set
of labeled data that is tagged with answers while receiving continuous feedback on the precision of its
solutions.

Supervised learning is useful for tasks that categorize (classification) or estimate the relationships
among variables (regression). Its applications include identifying suspicious activities in financial
systems and recognizing faces, objects, speech, gestures, or handwriting.

An example of supervised learning is neural networks, each consisting of an input layer, one or several
intermediate (or “hidden”) layers, and an output layer. Signals such as images, sounds, or texts are
2

converted into numbers at the input layer and then processed at the intermediate and the output
layers.

Convolutional and Recurrent Neural Networks are the most commonly seen neural networks.

A Convolutional Neural Network extracts features from an input signal, be it images or voice files,
while preserving the spatial relationship between the features for further analysis. Convolutional Neural
Networks are excellent in computer vision work, such as facial and speech recognition; as a result, it is
well-suited for operating self-driving cars, where image recognition is crucial in identifying other
vehicles, pedestrians, or road obstacles and alerting the self-driving vehicle of potential danger.

In a Recurrent Neural Network, the output of a processing layer is fed back to the same layer as input
for correction. If the prediction is wrong, the system self-learns and improves the prediction next time.
This type of neural network is highly effective for text-to-speech conversion. Recurrent Neural
Network is used principally for long, context-rich inputs, such as sentences that contain words with
double meanings, as in “fall” in “the fall season” and “fall down”, or audio files that contain different
words that have the same pronunciation, as in “their” and “there”.

Lastly, large, complex tasks may require a modular neural network, which consists of different
networks that function independently to perform sub-tasks; since the networks will not impede the
others, they will increase the overall computation speed of the task.

Transfer Learning

Supervised learning requires a large dataset that is completely labeled. However, assembling large,
complete datasets for every specific application is challenging and impractical.

Transfer learning deals with the shortage of specific, complete datasets by reusing the input and
middle layers of a model that has been trained with a dataset (the pre-trained model) so that it only
needs to retrain the final layers for the new task. The parameters of the pre-trained model will be used
in the beginning and then adjusted during the training to achieve maximal accuracy. Moreover, by
circumventing the need to train all the layers from scratch, transfer learning will significantly shorten
the overall training time for each specific application.

There are many pre-trained models, of which the most popular ones include the Mask R-CNN dataset
for object instance segmentation, YOLOv2 for object detection, VGG-Face Model for facial recognition,
Keras VGG16 model to classify tomatoes by their ripeness, and the Stanford car dataset for car
classification.

While transfer learning solves the lack of complete and unique datasets, it also has drawbacks, such
as being constrained by the pre-existing parameters in the pre-trained dataset, and it must proceed
slowly to prevent distortion.

Unsupervised Learning
3

In unsupervised learning, the algorithm tries to extract features from a set of unlabeled data, which
can be examples or signals with various attributes, to find the underlying patterns without any explicit
instruction.

As a result, unsupervised learning is useful in tasks that determine the association between features
or attributes by grouping (clustering), such as predicting what other products a customer might also
like based on his previous purchases.

Unsupervised learning can organize the data differently depending on the question one asks.
Therefore, asking the right question or asking a question the right way matters more in unsupervised
learning than in other types of learning.

Semi-supervised Learning

Supervised learning is most suitable for problems with many available reference points but not for
scenarios where labeling all the examples is impossible.

In semi-supervised learning, the algorithm trains with partially labeled datasets. In the case of
identifying tumors in CT scans or MRI images, having a trained radiologist label a small subset of
tumors will improve the algorithm’s accuracy over its unsupervised work significantly.

Ensemble learning

Ensemble learning combines multiple algorithms to achieve more accurate predictions than those by
any of the algorithms alone. One famous application of this method happened during the Netflix Prize
in 2006, where competing teams were given information on how half of the users in a dataset rated a
large number of movies and tasked to figure out how the other half of the users would rate the same
films. The winning team used the ensemble method to beat Netflix’s in-house algorithm.

Reinforcement Learning

Reinforcement learning continuously analyzes the cues from its environment to calculate to reach the
best next step. It sees applications mostly in control problems or games like chess or Go. For
example, in the late 1990s, the Deep Blue computer used reinforcement learning to fight world chess
champion Gary Kasparov to a draw, and in 2018, AlphaGo used the same method to beat Lee Sedol,
one of Go’s top players.

Machine Learning at the Cloud Versus the Edge

Traditionally, machine learning with industrial applications happens at a physical data center or the
virtual cloud, supported with sufficient processing capacity and electricity. But with the advent of the
Internet of Things (IoT), this model is facing challenges.
4

IoT devices away from the central cloud (hence “at the edge”) continuously collect a large amount of
data. Transferring these data to the central cloud for learning and then re-deploying to the edge is not
only expensive but also time-consuming; the associated time lag will make operations that require
real-time decision-making (or inference), such as in autonomous vehicles or military drones,
impossible. Also, data transfer may pose a threat to data security and integrity.
One way to solve this problem is to have machine learning take place at the edge. However, this
model also has drawbacks. For example, IoT devices are usually powered by small batteries and
installed in locations that make energy replenishment difficult, if not impossible; thus, energy supply
for machine learning will be an issue. In addition, the processing power provided by IoT devices may
be insufficient for machine learning. Therefore, there needs to be hardware improvement if machine
learning at the edge is to happen.

Part 2 of this series will discuss the hardware requirement for industrial and IoT edge machine
learning.