1. Machine Learning Basics

Core concepts that underpin all other topics.

• Types of Learning:
  • Supervised: Learn mapping from inputs to labeled outputs (e.g., classification, regression).
  • Unsupervised: Find hidden structure in unlabeled data (e.g., clustering, dimensionality reduction).
  • Reinforcement Learning: Learn actions via rewards/punishments.
• Key Ideas:
  • Training / validation / test sets.
  • Bias-variance tradeoff (underfitting ↔ overfitting).
  • Loss functions (MSE, cross-entropy).
  • Optimization: Gradient Descent (SGD, Adam).
  • Regularization: L1/L2, dropout, early stopping.

2. Deep Learning & Neural Networks

Scaling up ML with multi-layer, differentiable functions.

• Basic Unit: Neuron → weighted sum + non-linear activation (ReLU, sigmoid, tanh).
• Key Architectures:
  • FNN (Feedforward): Fully connected layers.
  • CNN: Spatial hierarchies via convolution + pooling (image, video).
  • RNN/LSTM: Sequential data (text, time series) — now largely replaced by Transformers.
• Training Deep Nets:
  • Backpropagation + automatic differentiation.
  • Vanishing/exploding gradients → skip connections, batch norm, residual nets (ResNet).
• Practical tricks:
  • Weight initialization (Xavier/He).
  • Learning rate scheduling.
  • Data augmentation.

3. Transformers & LLMs

Current dominant paradigm for sequence modeling, powering large language models.

• Self-Attention (core innovation):
  • Query, Key, Value matrices.
  • Scaled dot-product attention: $\text{Attention}(Q,K,V)=\text{softmax}\left(\frac{Q{K}^T}{\sqrt{d_k}}\right)V$Attention(Q,K,V)=softmax(dk​​QKT​)V.
  • Allows each token to directly attend to all others → solves long-range dependency limits of RNNs.
• Transformer Block:
  • Multi-head attention (parallel attention heads).
  • Feedforward network (per token).
  • Residual connections + layer norm.
  • Positional encodings (sinusoidal or learned).
• LLM Evolution:
  • Encoder-only (BERT) → understanding tasks.
  • Decoder-only (GPT) → generation, few-shot learning.
  • Encoder-decoder (T5, BART) → translation, summarization.
• Scaling Laws: Performance improves predictably with data, parameters, compute.
• Modern LLM concepts:
  • Pretraining (next-token prediction on web-scale text) + fine-tuning (SFT, RLHF).
  • In-context learning, Chain-of-Thought.
  • Efficient variants: LLaMA, Mistral; MoE (Mixture of Experts) like Mixtral, GPT-4.

4. Generative Models Deep Dive

Models that learn the data distribution $p(x)$p(x) to generate new samples.

Taxonomy & Key Methods:

Current State of the Art:

• Text-to-image: Diffusion models (Stable Diffusion, DALL-E 3, Midjourney, Flux).
• Text-to-video: Sora (diffusion transformer), Runway Gen-2.
• Text generation: Autoregressive LLMs (GPT-4, Claude, Gemini).
• Multimodal generation: Combined diffusion + LLM (e.g., any-to-any models).

Key Insight (2025+): Diffusion + Transformer backbones → dominant for high-quality continuous generation (image, video, audio). LLMs + discrete tokens for text/code. Hybrid models emerging.

Summary Comparison Table

2. Machine Learning Basics

Machine Learning (ML) is a subset of AI where systems learn patterns from data instead of being explicitly programmed.

Core Types of Machine Learning

Key ML Algorithms (Classical)

• Regression: Linear Regression, Polynomial Regression
• Classification: Logistic Regression, Decision Trees, Random Forest, SVM, Naive Bayes, k-NN
• Clustering: K-Means, Hierarchical, DBSCAN
• Dimensionality Reduction: PCA, t-SNE, UMAP
• Ensemble Methods: Bagging, Boosting (XGBoost, LightGBM, CatBoost)

Fundamental Concepts

• Bias-Variance Tradeoff: High bias = underfitting, High variance = overfitting.
• Overfitting vs Underfitting:
  • Overfitting: Model memorizes training data, poor on new data.
  • Underfitting: Model too simple to capture patterns.
• Evaluation Metrics:
  • Regression: MSE, RMSE, MAE, R²
  • Classification: Accuracy, Precision, Recall, F1-Score, ROC-AUC, Confusion Matrix
• Cross-Validation: k-fold CV to get reliable performance estimate.
• Feature Engineering: Creating better input features (still very important).

Training Process:

1. Split data → Train / Validation / Test sets
2. Choose model + hyperparameters
3. Train on training set
4. Tune on validation set
5. Evaluate on unseen test set

3. Deep Learning & Neural Networks

Deep Learning is Machine Learning using artificial neural networks with many layers (hence “deep”).

Biological Inspiration

• Biological neuron → Artificial Neuron (Perceptron)

Core Components of a Neural Network

1. Neurons (Nodes)
2. Layers:
   • Input Layer
   • Hidden Layers (this is where depth comes from)
   • Output Layer
3. Weights & Biases (learnable parameters)
4. Activation Functions (introduce non-linearity):Common ones:
   • ReLU (Rectified Linear Unit): f(x) = max(0, x) — most popular
   • Sigmoid: 1 / (1 + e^(-x))
   • Tanh
   • Leaky ReLU, GELU, Swish (modern)

Forward Propagation

Input → Weighted sum → Activation → Next layer → Final output

Backpropagation (The Learning Algorithm)

• Calculate error (Loss function)
• Compute gradients using Chain Rule
• Update weights using Gradient Descent (or variants: SGD, Adam, RMSprop)

Loss Functions:

• Mean Squared Error (regression)
• Cross-Entropy (classification)
• Binary Cross-Entropy

Popular Architectures

• Feedforward Neural Networks (MLP) — Basic
• Convolutional Neural Networks (CNNs) — Best for images & spatial data
• Recurrent Neural Networks (RNNs) — For sequences (old)
• LSTMs / GRUs — Improved RNNs (better at long dependencies)
• Transformers — Current dominant architecture (next section)

Why Deep Learning Works Well:

• Automatic feature learning (no manual feature engineering needed)
• Hierarchical representations (edges → shapes → objects)

Challenges:

• Requires massive data
• Computationally expensive (needs GPUs/TPUs)
• Black-box nature (hard to interpret)

4. Transformers & Large Language Models (LLMs)

The Transformer (introduced in 2017 paper “Attention Is All You Need”) is the most important architecture in modern AI.

Key Innovation: Self-Attention Mechanism

Instead of processing sequentially (like RNNs), Transformers process entire sequences in parallel using attention.

Formula (Simplified):

$$\text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V$$Attention(Q,K,V)=softmax(dk​​QKT​)V

• Q = Query, K = Key, V = Value
• Scaled dot-product attention

Transformer Architecture

• Encoder (for understanding) — Used in BERT
• Decoder (for generation) — Used in GPT models
• Encoder-Decoder — Used in translation (T5, BART)

Components:

• Multi-Head Self-Attention
• Feed-Forward Networks
• Layer Normalization + Residual Connections
• Positional Encoding (since attention has no sense of order)

Large Language Models (LLMs)

Modern LLMs are Decoder-only Transformers trained on massive text data.

Training Objective: Next Token Prediction (Causal Language Modeling)

Given “The sky is”, predict next word “blue”.

Key Scaling:

• More parameters → Better performance (Emergent abilities appear)
• More data
• More compute

Major LLM Families (as of 2026):

Techniques Used in LLMs:

• Pre-training (massive unsupervised data)
• Instruction Tuning / Supervised Fine-Tuning (SFT)
• RLHF (Reinforcement Learning from Human Feedback)
• Chain-of-Thought (CoT) prompting
• Test-time Compute (o1-style reasoning)

5. Generative Models Deep Dive

Generative Models learn the underlying probability distribution of data to create new samples.

Major Types

1. Generative Adversarial Networks (GANs)
   • Generator vs Discriminator game
   • Excellent image quality but mode collapse issues
   • Variants: StyleGAN, CycleGAN, BigGAN
2. Variational Autoencoders (VAEs)
   • Encoder compresses data into latent distribution
   • Decoder generates from latent space
   • Good for smooth interpolation
3. Flow-based Models (Normalizing Flows)
   • Reversible transformations
   • Exact likelihood computation
4. Diffusion Models (Current King for Images)
   • Forward process: Gradually add Gaussian noise
   • Reverse process: Learn to denoise step-by-step
   • Models: Stable Diffusion, DALL·E 3, Midjourney v6, Flux, Imagen 3
5. Autoregressive Models
   • GPT-style: Generate one token at a time
   • Excellent for text, also used in image (PixelRNN, Parti)
6. Multimodal Generative Models
   • Generate across modalities (text → image, image → video, audio, etc.)

Comparison of Generative Models

Current Trends (2026):

• Consistency Models / Flow Matching — Faster diffusion
• Mixture of Experts (MoE) — Efficient scaling
• World Models & Video generation (Sora-like models)
• Agentic Generation — Models that plan before generating