13

What is AI?

Intelligence means the ability to process information

in order to make a decision.

CMPS 497/CMPE 471 Special Topics in Deep Learning 2

 What is Machine Learning?

CMPS 497/CMPE 471 Special Topics in Deep Learning 4

 Rules
Answers
Data

Answers
Rules
Data
 What is Deep Neural Network?

What is Deep Learning?

CMPS 497/CMPE 471 Special Topics in Deep Learning 7

 What is Deep Learning?

ARTIFICIAL
INTELLIGENCE
MACHINE LEARNING
Any technique that enables
Ability to leam without
computers to mimic
e><plicitJy being programmed
human behavior

UII
Teaching computers how to learn a task directly from raw data
5
 6
Figures from http://udlbook.com
 8
Figures from http://udlbook.com
 AI is all about deep learning
 Yes
 No

…...... is fitting mathematical models to observed data

 AI
 Machine Learning
 Deep Learning
CMPS 497/CMPE 471 Special Topics in Deep Learning 10
 Deep learning == Machine learning?
 Yes
 No

…...... is a type of machine learning

 Deep Neural Networks
 Supervised Learning
 Deep Learning
CMPS 497/CMPE 471 Special Topics in Deep Learning 11
12
 Supervised learning
• Define a mapping from input to output
• Learn this mapping from paired input/output (labeled) data examples

Model Model
Real world input Real world output
input output

CMPS 497/CMPE 471 Special Topics in Deep Learning 13

Simple example …
• Predict the height of a child given his/her age.

Model Model
Age of child Height of child
input output

CMPS 497/CMPE 471 Special Topics in Deep Learning 14

What is a supervised learning model?

Regression

• An equation relating input (age) to output (height)

• Search through family of possible equations to find one that fits training data well
CMPS 497/CMPE 471 Special Topics in Deep Learning 15
Figures from http://udlbook.com
What is a supervised learning model?

Deep neural networks are just a very flexible family of equations

Fitting deep neural networks = “Deep Learning”
CMPS 497/CMPE 471 Special Topics in Deep Learning 16
Figures from http://udlbook.com
Regression

Deep learning
model

Univariate regression problem (one output, real value)

Fully connected network
CMPS 497/CMPE 471 Special Topics in Deep Learning 17
Figures from http://udlbook.com
Graph regression

Deep learning
model

Multivariate regression problem (>1 output, real value)

Graph neural network
CMPS 497/CMPE 471 Special Topics in Deep Learning 18
Figures from http://udlbook.com
Text classification

Deep learning
model

Binary classification problem (two discrete classes)

Transformer network
CMPS 497/CMPE 471 Special Topics in Deep Learning 19
Figures from http://udlbook.com
Music genre classification

Deep learning
model

Multiclass classification problem (discrete classes, >2 possible values)

Recurrent neural network (RNN)
CMPS 497/CMPE 471 Special Topics in Deep Learning 20
Figures from http://udlbook.com
Image classification

Deep learning
model

Multiclass classification problem (discrete classes, >2 possible values)

Convolutional network
CMPS 497/CMPE 471 Special Topics in Deep Learning 21
Figures from http://udlbook.com
 Image segmentation

Multivariate binary classification problem (many outputs, 2 discrete classes)

Convolutional encoder-decoder network
CMPS 497/CMPE 471 Special Topics in Deep Learning 22
Figures from http://udlbook.com
Depth estimation

Multivariate regression problem (many outputs, continuous)

Convolutional encoder-decoder network

CMPS 497/CMPE 471 Special Topics in Deep Learning 23
Figures from http://udlbook.com
Terms
• Regression = continuous numbers as output
• Classification = discrete classes as output
• Two class and multiclass classification treated differently
• Univariate = one output
• Multivariate = more than one output

CMPS 497/CMPE 471 Special Topics in Deep Learning 24

“Given a news article, I want to predict if it is political,
sports, or economical”.
What type of a problem is that?
 multivariate classification
 univariate regression
 binary classification
 multivariate regression
 multiclass classification
CMPS 497/CMPE 471 Special Topics in Deep Learning 26
 Other type of examples …

CMPS 497/CMPE 471 Special Topics in Deep Learning 27

Translation

Deep learning
model

CMPS 497/CMPE 471 Special Topics in Deep Learning 28

Figures from http://udlbook.com
Image captioning

Deep learning
model

CMPS 497/CMPE 471 Special Topics in Deep Learning 29

Figures from http://udlbook.com
Image generation from text

Deep learning
model

CMPS 497/CMPE 471 Special Topics in Deep Learning 30

Figures from http://udlbook.com
 What do these examples
have in common?

CMPS 497/CMPE 471 Special Topics in Deep Learning 31

What do these examples have in common?
• Very complex relationship between input and output
• Sometimes may be many possible valid answers
• But outputs (and sometimes inputs) obey rules

Language obeys grammatical rules Natural images also have “rules”

Learn the “grammar” of the data

from LOTS of unlabeled examples
CMPS 497/CMPE 471 Special Topics in Deep Learning 32
Figures from http://udlbook.com
33
Unsupervised Learning
• Learning from (input) data without (output) labels, i.e., no supervision
• Rather than learning mapping from input to output, learn the structure
of the data,

• Clustering
• Finding outliers
• Generating new examples
• Filling in missing data

CMPS 497/CMPE 471 Special Topics in Deep Learning 34

DeepCluster: Deep Clustering for Unsupervised Learning of Visual Features (Caron et al., 2018) 35
Figures from http://udlbook.com
Unsupervised Learning
• Learning about a dataset without labels
• e.g., clustering
• Generative models can create examples
• e.g., generative adversarial networks

CMPS 497/CMPE 471 Special Topics in Deep Learning 36

Unsupervised Learning
• Learning about a dataset without labels
• e.g., clustering
• Generative models can create examples
• e.g., generative adversarial networks
• PGMs learn distribution over data
• e.g., variational autoencoders,
• e.g., normalizing flows,
• e.g., diffusion models

CMPS 497/CMPE 471 Special Topics in Deep Learning 37

Figures from http://udlbook.com
Generative models

CMPS 497/CMPE 471 Special Topics in Deep Learning 38

Figures from http://udlbook.com
Generative models

CMPS 497/CMPE 471 Special Topics in Deep Learning 39

Figures from http://udlbook.com
Generative models

CMPS 497/CMPE 471 Special Topics in Deep Learning 40

Figures from http://udlbook.com
Conditional synthesis

Original image Removed regions Generated regions

CMPS 497/CMPE 471 Special Topics in Deep Learning 41

Figures from http://udlbook.com
Figures from http://udlbook.com 42
 ChatGPT
CMPS 497/CMPE 471 Special Topics in Deep Learning 43
CMPS 497/CMPE 471 Special Topics in Deep Learning 44
45
Reinforcement learning

• Build an agent which lives in a world and can perform certain actions
at each time step.
• Goal: take actions to change the state so that you receive rewards
• You don’t receive any data – you have to explore the environment
yourself to gather data as you go.
CMPS 497/CMPE 471 Special Topics in Deep Learning 46
Example: chess
• States are valid states of the chess board
• Actions at a given time are valid possible moves
• Positive rewards for taking pieces, negative rewards for losing them

CMPS 497/CMPE 471 Special Topics in Deep Learning 47

Figures from http://udlbook.com
Example: chess
• States are valid states of the chess board
• Actions at a given time are valid possible moves
• Positive rewards for taking pieces, negative rewards for losing them

CMPS 497/CMPE 471 Special Topics in Deep Learning 48

Figures from http://udlbook.com
Why is this difficult?
• Stochastic
• Make the same move twice, the opponent might not do the same thing
• Rewards also stochastic (opponent does or doesn’t take your piece)
• Temporal credit assignment problem
• Did we get the reward because of this move? Or because we made good
tactical decisions somewhere in the past?
• Exploration-exploitation trade-off
• If we found a good opening, should we use this?
• Or should we try other things, hoping for something better?

CMPS 497/CMPE 471 Special Topics in Deep Learning 49

Landmarks in (Machine) Deep Learning
• 1958 Perceptron (Simple `neural’ model)
• 1986 Backpropagation (Practical Deep Neural networks)
• 1989 Convolutional networks (Supervised learning)
• 2012 AlexNet Image classification (Supervised learning)
• 2014 Generative adversarial networks (Unsupervised learning)
• 2014 Deep Q-Learning -- Atari games (Reinforcement learning)
• 2016 AlphaGo (Reinforcement learning)
• 2017 Machine translation (Supervised learning)
• 2019 Language models ((Un)supervised learning)
• 2022 Dall-E2 Image synthesis from text prompts ((Un)supervised learning)
• 2022 ChatGPT ((Un)supervised learning)
• 2023 GPT4 Multimodal model ((Un)supervised learning)
CMPS 497/CMPE 471 Special Topics in Deep Learning 50
2018 Turing award winners

CMPS 497/CMPE 471 Special Topics in Deep Learning 51

Where are we going?
• Supervised learning (overview with regression example)
• Shallow neural networks (a more flexible model)
• Deep neural networks (an even more flexible model)
• Loss functions (guiding the training)
• How to train neural networks (gradient descent and variants)
• How to measure performance of neural networks (generalization)