EC 502

Artificial Neural Networks

(ANN)

EC 608 AML 1
 Meet Your Instructor
Dr. Paridhi Athe
Assistant Professor
Department of Electronics and Communication Engineering (ECE)
IIIT Bhopal

Academic Background
• Ph.D. in Nuclear Engineering, North Carolina State University (NCSU), USA
• [Link]. in Nuclear Engineering, IIT Kanpur
• [Link]. in Semiconductor Devices & VLSI Technology, IIT Roorkee
• B.E. in Electronics and Communication Engineering, SGSITS Indore

EC 608 AML 2
 Research Specialization & Recognitions
Area of specialization
• AI assisted design and analysis of advanced energy systems
• Autonomous Control Systems for Energy Applications
• Cyber-Physical Systems
• Risk-Informed Decision Analysis
• Digital Twins and Multimodal AI Integration
• Credibility and Trust in AI applications.

Contributed to different U.S. Department of Energy–funded projects as a Postdoctoral researcher and Research faculty at NC
State University with research published in reputed nuclear engineering journals and conferences

Awards & Recognitions

• American Nuclear Society, NURETH-20 Best Paper Award, 20th International Topical Meeting on Nuclear Reactor Thermal
Hydraulics (NURETH), Washington, D.C. (a premier international conference in nuclear reactor thermal hydraulic).

• 2023 American Nuclear Society Thermal Hydraulic Division Best Paper Award (awarded for the best paper of the year across
all conferences sponsored by the division).
EC 314 CONTROL SYSTEM 3
 Relevant Projects
Postdoc & Research faculty (NCSU)
• Uncertainty Quantification Approach for Digital Twin-based Autonomous Control (Dec 2023-Feb 2025)
• Funded by US Department of Energy via Idaho National Lab, Idaho Falls, USA
• Testing generative AI for Digital Instrumentation and Control design task (April 2024- Sept 2024)
• Funded by US Department of Energy via Idaho National Lab, Idaho Falls, USA
• Development of methodologies for assessment of predictive capability maturity of Modeling and Simulation (M & S) tools
used in advanced reactor applications (Sep 2021-Sept 2023)
• Funded by U.S. Department of Energy via Advanced Reactor Demonstration Program
• Development of a Nearly Autonomous Management and Control (NAMAC) system for advanced nuclear reactors (Oct 2018-
Sept 2021)
• Funded by US Department of Energy, Advanced Research Projects Agency-Energy
PhD (NCSU)
• Application of Predictive Capability Maturity Model for the assessment of Consortium for Advanced Simulation of Light
Water Reactor (CASL) Virtual Environment for Reactor Analysis (VERA) codes (Jan 2015-Jun 2018)
• Funded by US Department of Energy, Consortium for Advanced Simulation of Light Water Reactor
• A Framework for Predictive Capability Maturity Assessment of Computer Simulation Codes (PhD thesis) (Jan 2015-Jun
2018)
• Funded by US Department of Energy, Consortium for Advanced Simulation of Light Water Reactor

EC 608 AML 4
Course content (Institute Website)

EC 608 AML 5
 Key references
Deep Learning
– Ian Goodfellow, Yoshua Bengio, Aaron Courville
– MIT Press, 2016

Neural Networks and Learning Machines 3e

-Simon Haykin
-Pearson, 2016

Relevant papers, articles, repositories, etc.

EC 608 AML 6
 Tentative Lecture Format
• The session begins with a 25-minute lecture.
• Lecture content covers key concepts and theories.

• The remaining time will be dedicated to Quiz or problem-solving.

EC 608 AML 7
 This week
• Introduction
• AI and its subfields
• History of AI
• Classification of AI/ML algorithms

EC 608 AML 8
 Artificial Intelligence
What is Intelligence?
• Intelligence is the cognitive ability to learn, reason,
solve problems, and adapt to new situations.
• Human intelligence involves complex problem-
solving, emotional understanding, and language
comprehension.
• Artificial Intelligence (AI) seeks to replicate or
simulate these aspects using machines.

Meaning of Cognition: the mental action or process of acquiring knowledge and understanding through thought,
experience, and the senses.
EC 608 AML 9
AI and Its Subfields
Term Definition Key Focus Examples

Artificial Simulation of human

Intelligence intelligence by
(AI) Virtual assistants,
Artificial Intelligence machines to perform Broad problem-solving
recommendation
(AI) tasks like reasoning, & cognition
systems
learning, and decision-
Machine making.
Learning
(ML)
Subset of AI that uses
Machine Learning mathematical models to Pattern recognition from Spam detection, price
(ML) find patterns in large data prediction
Deep datasets.
learning (DL)

Subset of ML that uses

multi-layered neural
Neural networks & Image recognition,
Deep Learning (DL) networks to learn
feature learning language translation
Generative AI complex patterns from
unstructured data.
Subset of deep learning
that creates new content Content creation ChatGPT, DALL·E,
Generative AI like text, images, music, through pattern MusicGen, GitHub
or code based on generation Copilot
learned data.
EC 608 AML 10
 DIKW pyramid

EC 608
AML [Link] 11
 Historical Background (1/4)
Early Foundations
• 1940s: Emergence of computing ideas. Alan Turing laid the groundwork
with the concept of a "universal machine."
• 1950: Alan Turing proposed the Turing Test to assess machine
intelligence.
• 1956: The Dartmouth Conference marked the official birth of the AI
field.

EC 608 AML 12
 Turing Test
What is the Turing Test?
• Proposed by Alan Turing in 1950 to assess machine intelligence.
• A machine passes the test if a human cannot reliably tell whether they are interacting with a machine or a human during a
conversation.
How It Works:
• A human judge converses with two unseen participants (one human, one machine).
• If the judge cannot consistently identify the machine, it is said to have passed the test.
Key Points:
• Focuses on language and behavior, not internal mechanisms.
• Conducted via text-based communication to remove bias.
Limitations:
• Doesn’t measure true understanding or reasoning.
• Can be fooled by simple tricks or evasive answers.
• Doesn’t account for physical or perceptual intelligence.
Relevance Today:
• Early chatbots (e.g., ELIZA) tried to pass it.
• Modern language models like ChatGPT approach it more closely.
EC 608 AML 13
 Historical Background (2/4)
Symbolic AI Era (1950s–1980s)
• AI systems based on symbolic rules and logic.
• Focused on tasks like solving algebra problems or playing chess.
• Example: Expert systems like MYCIN (medical diagnosis).
• MYCIN was an early expert system developed in the 1970s at Stanford
University to diagnose and treat bacterial infections. It used a rule-based
approach, with approximately 500 production rules, to analyze patient data,
identify potential bacterial causes, and recommend appropriate antibiotic
treatments.
• Limitations: These systems could not handle uncertainty or learn from
data.

EC 608 AML 14
 What is Symbolic AI?
• Symbolic AI (also known as Good Old-Fashioned AI or GOFAI) uses
symbols and logic rules to represent knowledge and perform reasoning.
• Characteristics:
• Based on manually written rules ("if-then" logic).
• Reasoning is transparent and interpretable.
• Works well in structured, well-defined environments.
• Example: "If (X is a bird) AND (X cannot fly), then (X is a penguin)."

EC 608 AML 15
 What is Sub-symbolic AI?
• Sub-symbolic AI uses data-driven models where intelligence emerges
from statistical patterns rather than hand-coded rules.
• Often called "connectionist AI," especially with neural networks.
• Characteristics:
• Learns from examples (data).
• Uses numeric vectors instead of symbolic representations.
• Typically less interpretable (black-box).
• Examples: Deep learning for vision and language (e.g., CNNs,
transformers).

EC 608 AML 16
 Comparison of Symbolic and Sub-symbolic AI
Feature Symbolic AI Sub-symbolic AI (ML/DL)
Basis Rules and symbols Data and patterns
Interpretability High Often low (black-box models)
Flexibility Low (manually coded) High (learns from data)
Examples Expert systems, logic programs Neural networks, deep learning
Strengths Transparency, reasoning Generalization, adaptability
Weaknesses Rigid, brittle Opaque, data-hungry

EC 608 AML 17
 Historical Background (3/4)
Machine Learning Emerges (1980s–1990s)
• Transition from rule-based systems to algorithms that learn from data.
• Emergence of decision trees, support vector machines, and early neural
networks.
• Focus shifted to adaptability and statistical pattern recognition.
AI Winters
• Periods of disillusionment due to limited real-world results.
• Reduced funding and enthusiasm in the 1970s and late 1980s.

EC 608 AML 18
 Historical Background (4/4)
Modern Era (2000s–Present)
• Advances in hardware (GPUs), data availability, and deep learning.
• 2012: AlexNet's success in ImageNet revolutionized image
classification.
• 2016: DeepMind's AlphaGo beat a world champion in Go.
• 2020s: Generative AI (e.g., ChatGPT) shows abilities in language, code,
and art

EC 608 AML 19
 Recent development in AI
2012 2013 2014 2015
• AlexNet (Deep • Computer vision • Generative • Residual nets
CNN) switches to neural Adversarial (ResNets), RNN &
network, VAE Networks (GANs) LSTM

2019 2018 2017 2016

• GPT-2 & improved • GPT-1, BERT, • Transformer • AlphaGo (deep
generative models Graph NN architecture Reinforcement learning,
Monte Carlo tree search)

2020 2021 2022 2023-25 and onwards

• GPT-3 & Self • AlphaFold 2, • ChatGPT (GPT 3.5) • LLaMA (Meta), GPT-4 (OpenAI),
Supervised Learning DALL.E, GitHub & Stable Diffusion Gemini(Google), Multimodal AI,
Copilot Multi agent AI (Deep research)

CNN: Convolutional Neural Network, VAEs: Variational Autoencoders, GANs: Generative Adversarial Networks, RNN: Recurrent Neural Network, LSTM: Long Short-
ECRepresentations
Term Memory, GPT: Generative Pre-trained Transformer, BERT: Bidirectional Encoder 608 AML from Transformers 20
 AlexNet, a deep convolutional neural

Recent development in AI
network (CNN)→ breakthrough in
image recognition tasks

2012 2013 2014 2015

• AlexNet (Deep • Computer vision • Generative • Residual nets
CNN) switches to neural Adversarial (ResNets), RNN &
network, VAE Networks (GANs) LSTM

2019 2018 2017 2016

• GPT-2 & improved • GPT-1, BERT, Graph • Transformer • AlphaGo (deep
generative models NN architecture Reinforcement learning,
Monte Carlo tree search)

2020 2021 2022 2023-25 and onwards

• GPT-3 & Self • AlphaFold 2, • ChatGPT (GPT 3.5) • LLaMA (Meta), GPT-4 (OpenAI),
Supervised DALL.E, GitHub & Stable Diffusion Gemini(Google), Multimodal AI,
Learning Copilot Multi agent AI (Deep research)

CNN: Convolutional Neural Network, VAEs: Variational Autoencoders, GANs: Generative Adversarial Networks, RNN: Recurrent Neural Network, LSTM: Long Short-
ECRepresentations
Term Memory, GPT: Generative Pre-trained Transformer, BERT: Bidirectional Encoder 608 AML from Transformers 21
 VAE→ learns to represent and generate data (images
Recent development in AI and sound)→ learns a compressed representation of the
input data in a lower dimensional space (latent space)

2012 2013 2014 2015

• AlexNet (Deep • Computer vision • Generative • Residual nets
CNN) switches to neural Adversarial (ResNets), RNN &
network, VAE Networks (GANs) LSTM

2019 2018 2017 2016

• GPT-2 & improved • GPT-1, BERT, Graph • Transformer • AlphaGo (deep
generative models NN architecture Reinforcement learning,
Monte Carlo tree search)

2020 2021 2022 2023-25 and onwards

• GPT-3 & Self • AlphaFold 2, • ChatGPT (GPT 3.5) • LLaMA (Meta), GPT-4 (OpenAI),
Supervised DALL.E, GitHub & Stable Diffusion Gemini(Google), Multimodal AI,
Learning Copilot Multi agent AI (Deep research)

CNN: Convolutional Neural Network, VAEs: Variational Autoencoders, GANs: Generative Adversarial Networks, RNN: Recurrent Neural Network, LSTM: Long Short-
ECRepresentations
Term Memory, GPT: Generative Pre-trained Transformer, BERT: Bidirectional Encoder 608 AML from Transformers 22
 Recent development in AI GAN → generator –
discriminator in a
game-like setup

2012 2013 2014 2015

• AlexNet (Deep • Computer vision • Generative • Residual nets
CNN) switches to neural Adversarial (ResNets), RNN &
network, VAE Networks (GANs) LSTM

2019 2018 2017 2016

• GPT-2 & improved • GPT-1, BERT, Graph • Transformer • AlphaGo (deep
generative models NN architecture Reinforcement learning,
Monte Carlo tree search)

2020 2021 2022 2023-25 and onwards

• GPT-3 & Self • AlphaFold 2, • ChatGPT (GPT 3.5) • LLaMA (Meta), GPT-4 (OpenAI),
Supervised DALL.E, GitHub & Stable Diffusion Gemini(Google), Multimodal AI,
Learning Copilot Multi agent AI (Deep research)

CNN: Convolutional Neural Network, VAEs: Variational Autoencoders, GANs: Generative Adversarial Networks, RNN: Recurrent Neural Network, LSTM: Long Short-
ECRepresentations
Term Memory, GPT: Generative Pre-trained Transformer, BERT: Bidirectional Encoder 608 AML from Transformers 23
 ResNet: solved vanishing
Recent development in AI gradient issue enabling
training much deeper NNs

2012 2013 2014 2015

• AlexNet (Deep • Computer vision • Generative • Residual nets
CNN) switches to neural Adversarial (ResNets), RNN &
network, VAE Networks (GANs) LSTM

2019 2018 2017 2016

• GPT-2 & improved • GPT-1, BERT, Graph • Transformer • AlphaGo (deep
generative models NN architecture Reinforcement learning,
Monte Carlo tree search)

2020 2021 2022 2023-25 and onwards

• GPT-3 & Self • AlphaFold 2, • ChatGPT (GPT 3.5) • LLaMA (Meta), GPT-4 (OpenAI),
Supervised DALL.E, GitHub & Stable Diffusion Gemini(Google), Multimodal AI,
Learning Copilot Multi agent AI (Deep research)

CNN: Convolutional Neural Network, VAEs: Variational Autoencoders, GANs: Generative Adversarial Networks, RNN: Recurrent Neural Network, LSTM: Long Short-
ECRepresentations
Term Memory, GPT: Generative Pre-trained Transformer, BERT: Bidirectional Encoder 608 AML from Transformers 24
 Recent development in AI
2012 2013 2014 2015
• AlexNet (Deep • Computer vision • Generative • Residual nets
CNN) switches to neural Adversarial (ResNets), RNN &
network, VAE Networks (GANs) LSTM
Google’s AlphaGo defeated
the world champion of Go, Lee
Sedol.

2019 2018 2017 2016

• GPT-2 & improved • GPT-1, BERT, Graph • Transformer • AlphaGo (deep
generative models NN architecture Reinforcement learning,
Monte Carlo tree search)

2020 2021 2022 2023-25 and onwards

• GPT-3 & Self • AlphaFold 2, • ChatGPT (GPT 3.5) • LLaMA (Meta), GPT-4 (OpenAI),
Supervised DALL.E, GitHub & Stable Diffusion Gemini(Google), Multimodal AI,
Learning Copilot Multi agent AI (Deep research)

CNN: Convolutional Neural Network, VAEs: Variational Autoencoders, GANs: Generative Adversarial Networks, RNN: Recurrent Neural Network, LSTM: Long Short-
ECRepresentations
Term Memory, GPT: Generative Pre-trained Transformer, BERT: Bidirectional Encoder 608 AML from Transformers 25
 Recent development in AI
The Transformer architecture, introduced in 2017 by a team
2012 2013 2014 2015
of researchers at Google, represents a major breakthrough in
• AlexNet (Deep • Computer vision • Generative
neural network • Residual nets
design. Its self-attention mechanism
CNN) switches to neural Adversarial scalability in training
enabled unprecedented (ResNets), RNN &
deep models,
network, VAE Networks (GANs)
and this architectural LSTM
innovation laid the foundation for
today’s Large Language Model era.

2019 2018 2017 2016

• GPT-2 & improved • GPT-1, BERT, Graph • Transformer • AlphaGo (deep
generative models NN architecture Reinforcement learning,
Monte Carlo tree search)

LLMs’ Era
2020 2021 2022 2023-25 and onwards
• GPT-3 & Self • AlphaFold 2, • ChatGPT (GPT 3.5) • LLaMA (Meta), GPT-4 (OpenAI),
Supervised DALL.E, GitHub & Stable Diffusion Gemini(Google), Multimodal AI,
Learning Copilot Multi agent AI (Deep research)

CNN: Convolutional Neural Network, VAEs: Variational Autoencoders, GANs: Generative Adversarial Networks, RNN: Recurrent Neural Network, LSTM: Long Short-
ECRepresentations
Term Memory, GPT: Generative Pre-trained Transformer, BERT: Bidirectional Encoder 608 AML from Transformers 26
 Recent development in AI
2012 2013 2014 2015
• AlexNet (Deep • Computer vision • Generative • Residual nets
CNN) switches to neural Adversarial (ResNets), RNN &
network, VAE Networks (GANs) LSTM

2019 2018 2017 2016

• GPT-2 & improved •AlphaFold
GPT-1, BERT, Graph
2: based on transformer•architecture,
Transformer • AlphaGo (deep
generative models NN architecture Reinforcement learning,
solved decades-old protein folding problem
Monte Carlo tree search)
→predict protein’s 3D structure based on its 1D
amino acid sequence
DALL.E → image generation from textual description

2020 2021 2022 2023-25 and onwards

• GPT-3 & Self • AlphaFold 2, • ChatGPT (GPT 3.5) • LLaMA (Meta), GPT-4 (OpenAI),
Supervised DALL.E, GitHub & Stable Diffusion Gemini(Google), Multimodal AI,
Learning Copilot Multi agent AI (Deep research)

CNN: Convolutional Neural Network, VAEs: Variational Autoencoders, GANs: Generative Adversarial Networks, RNN: Recurrent Neural Network, LSTM: Long Short-
ECRepresentations
Term Memory, GPT: Generative Pre-trained Transformer, BERT: Bidirectional Encoder 608 AML from Transformers 27
 Recent development in AI
Stable diffusion (image generation): a latent text-to-image
2012 2013 2014 2015
diffusion model capable of generating photo-realistic images
• AlexNet (Deep • Computer vision • GenerativeStable diffusion is an extension
from text descriptions. • Residual nets
of the
CNN) switches to neural Adversarial
traditional diffusion (ResNets),
models, which work by iteratively RNN &
adding
network, VAE Networks
noise to images (GANs)
and then LSTM
reversing the process to recover the
data. It was designed to speed up this process by operating
not directly on the input images, but instead on a lower-
dimensional representation, or latent space, of them. In
addition, the diffusion process is modified by adding the
2019 2018 2017
transformer-embedded 2016
text prompt from the user to the
• GPT-2 & improved • GPT-1, BERT, Graph network, allowing it to guide the image• AlphaGo
• Transformer generation(deep
process
generative models NN architecture
throughout each iteration. Reinforcement learning,
Monte Carlo tree search)

2020 2021 2022 2023-25 and onwards

• GPT-3 & Self • AlphaFold 2, • ChatGPT (GPT 3.5) • LLaMA (Meta), GPT-4 (OpenAI),
Supervised DALL.E, GitHub & Stable Diffusion Gemini(Google), Multimodal AI,
Learning Copilot Multi agent AI (Deep research)

CNN: Convolutional Neural Network, VAEs: Variational Autoencoders, GANs: Generative Adversarial Networks, RNN: Recurrent Neural Network, LSTM: Long Short-
ECRepresentations
Term Memory, GPT: Generative Pre-trained Transformer, BERT: Bidirectional Encoder 608 AML from Transformers 28
 Classification of AI
By Capabilities
• Narrow AI (Weak AI):
• Performs a single task well.
• Example: Google Translate, facial recognition.
• General AI (Strong AI):
• Hypothetical machines with human-level cognitive abilities.
• Still under research.
• Superintelligent AI:
• A level beyond human intelligence.
• Subject of philosophical and ethical debate.

EC 608 AML 29
 Classification of AI
By Functionalities
• Reactive Machines:
• No memory or past learning.
• Example: IBM's Deep Blue (chess-playing computer).
• Limited Memory:
• Uses past data to make decisions.
• Example: Self-driving cars tracking movement.
• Theory of Mind:
• Understands emotions and thoughts.
• Still theoretical; research ongoing.
• Self-aware AI:
• AI with consciousness and self-awareness.
• Currently science fiction.

EC 608 AML 30
Theory of Mind
Aspect Explanation

Theory of Mind (ToM) is the ability to attribute mental states -beliefs,

Definition intents, desires, emotions - to oneself and others.

Develops in children around age 4-5; allows understanding that others

Origin (Human Psychology) have different beliefs, knowledge, or intentions.

Enable machines to recognize and interpret human mental states to

Goal in AI guide adaptive, context-aware behavior.
Crucial for empathy, social interaction, collaborative problem-solving,
Why It's Important and trust in human-AI interaction.
No true ToM exists in AI today; current systems simulate basic
Current Status emotion or intent recognition.
Affective computing, multi-agent systems, cognitive modeling,
Examples of Related Research symbolic and subsymbolic integration.
Understanding implicit context, generalization, ethical issues (privacy,
Key Challenges manipulation), interpretability.
EC 608 AML 31
 Classification of ML
Type Description Examples
Learns from labeled input-output Email spam detection, credit
Supervised Learning
pairs scoring
Identifies patterns in data without Customer segmentation, topic
Unsupervised Learning
labels modeling
Combines small labeled and large
Semi-Supervised Learning Medical image analysis
unlabeled data
Learns by interacting with Robotics, game playing (e.g.,
Reinforcement Learning
environment via rewards AlphaZero)

EC 608 AML 32
Classification of ML
Machine
Learning

Supervised Unsupervised Semi- Reinforcement Self supervised

learning learning supervised learning learning

Regression

Classification

EC 608 AML 33
Classification of ML
Machine
Learning

Supervised Unsupervised Semi- Reinforcement Self supervised

learning learning supervised learning learning

Regression
• Teaches through labeled data
• Known input and output

Classification

EC 608 AML 34
Classification of ML
Machine
Learning

Supervised Unsupervised Reinforcement Self supervised

learning learning learning learning

Regression

• Teaches the machine to classify data.

Classification
• Predict discrete response (yes or no, distinct labels, etc.)
• Data tagging, categorization or separation into groups.
EC 608 AML 35
 Classification
• Classification is a type of supervised learning where the
goal is to assign a label (or class) to an input based on its
features.

• In the example shown, the model learns to separate

Class A (blue dots) and Class B (green triangles) using a
decision boundary.

• The decision boundary (black line) is learned from

labeled training data and is used to classify new, unseen
points.

• Common classification algorithms include logistic

regression, decision trees, support vector machines, and
neural networks.

EC 608 AML 36
Classification of ML
Machine
Learning

Supervised Unsupervised Semi- Reinforcement Self supervised

learning learning supervised learning learning

Regression • Teaches the machine to predict the future output as new data comes
• Predicts a number

Classification

EC 608 AML 37
 Regression
• Regression is a type of supervised learning used to
predict continuous numerical values.
• The goal is to learn a function 𝑓(𝑥) that maps input
features 𝑥 to a real-valued output 𝑦.
• Given examples (𝑥1 , 𝑦1 ),(𝑥2 , 𝑦2 ),…,(𝑥𝑛 , 𝑦𝑛 ), the
model learns to estimate 𝑦 for new inputs.
• Common real-world applications:
• Predicting house prices based on size, location,
etc.
• Forecasting temperature, stock prices, or
electricity demand.
• Estimating student grades, equipment lifespan,
etc.
• Typical algorithms used for regression: Linear
regression, Polynomial regression, Decision trees
and random forests, Neural networks (for complex,
non-linear patterns). EC 608 AML 38
Classification of ML
Machine
Learning

Supervised Unsupervised Semi- Reinforcement Self supervised

learning learning supervised learning learning

Regression • Groups similar data points together based on patterns in the data
• Does not need labeled data
• Identifies patterns in the data
• E.g., Clustering

Classification

EC 608 AML 39
 Clustering
• Clustering is a type of unsupervised learning where
the goal is to group similar data points together
based on patterns in the data.
• Unlike classification, clustering does not use labeled
data. The algorithm identifies natural groupings
(clusters) in the input.
• Real-world examples of clustering: Y
• Customer segmentation in marketing
• Document/topic grouping
• Anomaly detection (outliers are far from
clusters)
• Common clustering algorithms:
• K-Means
• Hierarchical Clustering
• Density-Based Spatial Clustering of
X
Applications with Noise
EC 608 AML 40
Classification of ML
Machine
Learning

Supervised Unsupervised Semi- Reinforcement Self supervised

learning learning supervised learning learning

Regression • Uses small labeled + large unlabeled dataset

• Bridges supervised and unsupervised learning
• Reduces labeling cost while improving accuracy
• Learns from labeled examples and generalizes using
unlabeled data
Classification • Example: classify emails with few labeled as spam/not
spam, rest unlabeled
EC 608 AML 41
Classification of ML
Machine
Learning

Supervised Unsupervised Semi- Reinforcement Self supervised

learning learning supervised learning learning

Regression • Agent–Environment interaction framework

• Learns by trial and error
• Uses rewards and penalties as feedback
• Goal: maximize cumulative reward over time
• Exploration vs. exploitation trade-off
Classification • Applications: robotics, game AI, recommendation systems,
autonomous driving

EC 608 AML 42
RL basic concept
▪ At each step the agent executes action, observe new state and receive the reward

[Link]
pinningup/rl_intro.html#part-1-key-
concepts-in-rl [Link]
43
Classification of ML
Machine
Learning

Supervised Unsupervised Semi- Reinforcement Self supervised

learning learning supervised learning learning

• Learns from unlabeled data

Regression • Generates labels automatically from the input
• Uses pretext tasks (e.g., predict missing words or image
parts)
• Scales to massive datasets without manual annotation
Classification • Produces general features for many downstream tasks
• LLMs (e.g., GPT) use it via next-token prediction

EC 608 AML 44
 Semi-Supervised vs. Self-Supervised Learning
Aspect Semi-Supervised Learning Self-Supervised Learning
Data Mix of labeled and unlabeled data Only unlabeled data
Generated automatically from the
Labels Provided for a small portion of data
data itself
Improve accuracy using limited Learn general representations from
Goal
labeled examples raw data
Classification with scarce labeled Pretraining models for various
Typical Use Case
data downstream tasks
Image classification with 1% Predict missing words in text (e.g.,
Example
labeled data GPT pretraining)

EC 608 AML 45
 Application of AI/ML
Domain AI/ML Applications Benefits

Disease diagnosis, drug discovery,

Healthcare Early detection, improved outcomes
personalized treatment

Fraud detection, credit scoring,

Finance Risk reduction, financial forecasting
algorithmic trading
Autonomous vehicles, traffic
Transportation Safer travel, efficient route planning
prediction, logistics optimization
Content recommendation, game AI, Personalized experiences, creative
Entertainment
content generation tools
Customer segmentation, demand Better targeting, improved customer
Retail
forecasting, chatbots service
Chatbots, language translation, Real-time communication, market
NLP
sentiment analysis insights
Predictive maintenance, quality
Manufacturing
control, robotic automation

EC 608 AML 46
Mathematical Framework of ML Models

47
Classification Regression

Clustering

EC 608 AML 48
 Generalization, overfitting, capacity
• Learning is the process of searching a hypothesis space to find a model that fits the training data.
• A model overfits when it performs well on training data but poorly on unseen/test data.
• Generalization refers to how well the trained model performs on new, unseen data.
• The balance between underfitting and overfitting depends on the capacity (flexibility) of the model.

49
Underfitting or too low capacity (left), appropriate capacity (center) and overfitting or too high capacity (right) in a model
Curse of Dimensionality

Goodfellow, Ian, Yoshua Bengio, Aaron Courville, and

Yoshua Bengio. Deep learning. Vol. 1, no. 2.
Cambridge: MIT press, 2016.

EC 608 AML 50