Skip to content

CO3519 Artificial Intelligence
[[CO3519 Lecture 10 - Reinforcement Learning]]

Lecture DocumentsΒΆ

CO3519 Lecture 11.pdf

Written NotesΒΆ

  • No written notes made all content from the lecture.

Course Overview (Recap)ΒΆ

  • Module: Artificial Intelligence
  • Course ID: CO3519
  • Number of Assessments: 2
  • 1st Assignment - Weightage => 50%, Contents => Facial Expression Recognition using ML. βœ”οΈ
  • 2nd Assignment - Weightage => 50%, Contents => Implementation of Deep Learning Algorithm and its visual demonstration.
  • Pre-Requisites: Python, Advanced Programming, Basic Machine Learning
  • Study/Support Material: It will be provided along each lecture on blackboard.

Learning ObjectivesΒΆ

  • Brief recap of traditional AI and ML concepts
  • History of advanced AI (Deep Learning)
  • What is DL
  • What is perceptron

This Semester CoversΒΆ

How LLM work and CNN (conventional neural networks).

Traditional AI and ML RecapΒΆ

Key Points Recap:
- AI - Simulating Human Intelligence (rule-based systems, expert systems)
- Machine Learning - Learning patterns from data (Linear Regression, Decision Trees)

Limitations of Traditional ML:
- Feature Engineering - Handcrafting Features is tedious.
- Performance - Struggles with high-dimensional data like images or text.

Why Deep Learning?ΒΆ

Basic Steps in MLΒΆ

1) Collect data and extract features
2) Build Modal
3) Optimisation - Mimise Error

DL: Requires large amounts of data
ML: Can be effective with smaller, structured datasets

DL: can outperform machine learning for complex tasks like image recognition and natural language understanding
ML: can be more efficient and easier to implement for simplier tasks.

What is Deep LearningΒΆ

Sub-field of ML which attempts to learn high-level abstractions in data by utilising hierarchical architectures.

Using a neural network with several layers of nodes between input and output.

The series of layers between input and output do feature identification and processing in a series of stages just as a human brain does.

graph LR
    A[Input: Dog/Cat Image] --> B[Layer 1: Edges & Lines]
    B --> C[Layer 2: Shapes & Textures]
    C --> D[Layer 3: Eyes, Ears & Snouts]
    D --> E[Output: Prediction 'Dog']

    style A fill:#f9f,stroke:#333,stroke-width:2px
    style E fill:#9f9,stroke:#333,stroke-width:2px

History of DL and Neural NetworkΒΆ

timeline
    title History of Deep Learning & Neural Networks
    section Early Foundations
        1958 : Perceptron <br> (Rosenblatt)
        1960 : Adaline <br> (Widrow & Hoff)
        1969 : Perceptrons Book <br> (Minsky & Papert)
        1970 : Backpropagation <br> (Linnainmaa)
    section The AI Winter & Thaw
        1974 : Backprop Applied <br> (Werbos)
        1986 : Backprop Popularized <br> (Rumelhart, Hinton, Williams)
        1997 : LSTM <br> (Hochreiter & Schmidhuber)
        1998 : LeNet / OCR <br> (LeCun et al.)
    section Modern Renaissance
        2006 : Deep Learning Term Coined <br> (Hinton et al.)
        2009 : ImageNet Dataset <br> (Deng et al.)
        2012 : AlexNet <br> (Krizhevsky, Sutskever, Hinton)
        2015 : ResNet <br> (He et al.)
    section Advanced DL
        2016 : AlphaGo <br> (DeepMind)
        2017 : Transformers <br> (Vaswani et al.)
        2018 : BERT <br> (Google)
    section Recent Developments
        2021 : DALL-E & Codex <br> (OpenAI)
        2022 : AlphaTensor <br> (DeepMind)
        2023 : GPT-4, SAM, CLIP <br> (Generative AI Boom)
ΒΆ

PerceptronΒΆ

Perceptrons, Rosenblatt 1957/58

  • Developed first implementation of perceptron
  • Machine consist of 400 photocells, classify images
  • Perceptron is a one-layer netowkr that an classify things into two parts.

A simple Neural NetworkΒΆ

```mermaid diagram
graph LR
subgraph Inputs
x1((x1))
x2((x2))
x3((x3))
end

subgraph Weights
w1(w1)
w2(w2)
w3(w3)
end

subgraph Neuron
S["Weighted Sum <br/> Ξ£ (x*w) + b"]
A["Activation Function <br/> (Step/Sign)"]
end

y((Output y))

%% FIX: Use standard arrows or add spaces like -- "text" -->
x1 --> S
x2 --> S
x3 --> S

S --> A
A --> y

style S fill:#f9f,stroke:#333
style A fill:#ff9,stroke:#333

```

Biological InspirationΒΆ

  • Neurons: Inspired by biological neurons in the brain.
  • Synapses: The "Weights" (\(w\)) mimic the synaptic strength between neurons. A higher weight means a stronger connection.

  • All-or-None Law: Just like a biological neuron either fires an action potential or doesn't, the Perceptron uses an activation function to output a crisp 1 (Fire) or 0 (Don't Fire) based on whether the signal crosses a threshold.

  • These neurons receive signals from other neurons

  • Depending on the strength, the neuron can either active or off/stay.

A Perceptron ExampleΒΆ

$\(z = (x_1 \cdot w_1) + (x_2 \cdot w_2) + (x_3 \cdot w_3) + b\)$
x (inputs) - data e.g. water temp.
w (weights) - importance of that data
b (bias) - shifting threshold
z (net input) - total strength of the signal before decision made

Next WeekΒΆ

  • Convolutional Neural Networks (CNNs) - Basics
  • Introduction To CNNs - Why CNNs are effective for image data
  • Key Components:
  • Convolution Layers
  • Pooling Layers
  • Filters
  • Stride
  • Padding
  • Activation Functions
  • Loss functions
  • Gradient descent

CO3519 Lecture 12 - Basics of Deep Learning