Scribd
Upload
77 views

Deep Learning: Hoàng Huy Minh Hoàng Thảo Lan Chi Phạm Huy Thiên Phúc Trương Huỳnh Đăng Khoa

Deep learning is a form of machine learning using neural networks with multiple layers between input and output. [1] Recent advances in deep learning include generating photorealistic videos from semantic labels or poses using generative adversarial networks, and language models like BERT that achieve state-of-the-art results on natural language tasks by considering both left and right context of words. [2] Deep learning has many applications from image recognition to machine translation and is an active area of research.

Uploaded by

sandy milk min
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
77 views

Deep Learning: Hoàng Huy Minh Hoàng Thảo Lan Chi Phạm Huy Thiên Phúc Trương Huỳnh Đăng Khoa

Deep learning is a form of machine learning using neural networks with multiple layers between input and output. [1] Recent advances in deep learning include generating photorealistic videos from semantic labels or poses using generative adversarial networks, and language models like BERT that achieve state-of-the-art results on natural language tasks by considering both left and right context of words. [2] Deep learning has many applications from image recognition to machine translation and is an active area of research.

Uploaded by

sandy milk min
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 25

Deep Learning

CC03
Hoàng Huy Minh
Hoàng Thảo Lan Chi
Phạm Huy Thiên Phúc
Trương Huỳnh Đăng Khoa
What is
deep learning?
Deep learning is a subset of machine learning, which is a
neural network with three or more layers
3
What is deep leaning?

➔ These neural networks try to copy how our brains work so it can “learn”
from large amounts of data that is unstructured or unlabelled.
➔ While a neural network with 1 layer still work, additional hidden layers
increase accuracy

4
Applications of deep learning

➔ Deep learning can improve automation, performing analytical and


physical tasks without human intervention.
➔ Examples:

Credit card fraud detection Voice-enable TV remotes Self-driving car


Digital assistants

5
History
and Development
6
7
Prominant
DL architectures

8
Traning with example
sets

9
1. Convolutional neural networks (CNN)
Structure Applications
• a multilayer neural network • image processing
• early layers recognize features and later layers • video recognition
recombine them into higher-level attributes of the • natural language
input processing
• Training: using back-propagation

10
2. Recurrent neural network (RNN)
Structure
• Connections to the same layer or prior layers that
help maintain memory of past inputs and model
problems in time
• Training: standard back-propagation or BPTT

Applications
• speech recognition
• handwriting recognition

11
3. LSTM networks

Structure Application
• memory cells Image and video captioning systems
• 3 gates: input, forget,
output, and weights to
control each gate
• Training: BPTT (Back-
propagation in time)

12
4. Gated recurrent unit (GRU) networks

Structure
• Simplified version of LSTM, more efficient,
less expressive
• 2 gates: update gate and reset gate
• Training: similar to LSTM

Applications
• text compression
• handwriting recognition
• speech recognition
• gesture recognition
• image captioning 13
14
1. Self-organized maps
Structure
• Clusters of the input data set
• Random weights are initialized to each feature of the input record and represent the input
node
• Nodes with the least distance between input and output node is most accurate and become
center points

Applications
• Dimensionality reduction
• Radiant grade result
• Cluster visualization

15
2. Autoencoders
Structure
• 3 layers: input, hidden, and output
layers
• Hidden layers have less nodes than
input
• Output layer decodes to
reconstructs input
• The weights are adjusted to
minimize the error
Applications
• Learn continuously
• Dimensionality reduction
• Data interpolation
• Data compression/decompression. 16
3. Restricted Boltzmann Machines (RBM)
Structure
• 2 layers: input and hidden layers
• Nodes are connected between 2 layers but not
within
• Neurons are activated at random
• Calculate probability distribution of the training
set by stochastic approach
• Hidden and visible bias

Applications
• Dimensionality reduction
• Collaborative filtering
17
Outstanding
achievements

18
Video to video synthesis
Ting-Chun Wang 2018
Architecture
Aim
Produce a high-resolution,
photorealistic video output from a
segmented input video in a diverse
set of formats

Method
1 generator to create images
2 discriminator to ensure the quality
of output videos

19
Video to video synthesis
1. Semantic Labels → Cityscapes Street Views

 Semantic labels are used for editing


and creating content
 Multiple results from the same
inputs
 Meaning of the labels can be
changed

20
Video to video synthesis
2. Edge -> Face 3. Pose -> Body
 Create videos of a human face
 Create realistic dance videos from
from edge map video
video of poses
 Generate different faces from
 Can change the clothing and give
the same input
out consistent shadows
 Can change the facial
appearance

21
Video to video synthesis
4. Frame prediction
2 tasks:
 Synthesize future masks from observed
frames
 Convert masks into videos by vid2vid
synthesis

22
Language models: BERT
Google AI Language team, 2019

BERT (Bidirectional Encoder Representations from Transformers)


• Context of a word is considered on both left and right instead of in 1 direction
• Show best results for 11 complex NLP tasks

23
Training of Google’s BERT

1. Mask Language Model (MLM)


• Mask 15% of input words
• The model predict the original word base
on other words in the sequence

2. Next Sentence Prediction (NSP)


• Learn relationships between sentences
• Receive pairs of sentences and learn to
predict if the 2nd sentence is the
subsequent sentence

24
Thanks for
listening!
Any questions?

25

You might also like