Deep Learning

Purpose

These notes treat deep learning as a modeling discipline, not just a list of branded architectures. The section starts with neural networks from first principles, then moves to the design choices that matter before the first training run, then covers the main architecture families that recur across language, vision, sequence modeling, and generative modeling.

The goal is not only to say what each model is. The goal is to make the implementation and the math line up. Most notes include:

• the forward equations
• the backpropagation view or training objective
• a small NumPy implementation that exposes the mechanics
• a more idiomatic PyTorch implementation, often with einops for shape clarity

Reading Path

1. Neural Networks from Scratch
2. Modeling, Architecture, and Data
3. Recurrent Neural Networks
4. Convolutional Neural Networks
5. Decoder-Only Transformers
6. Encoder-Decoder Transformers
7. Diffusion Models
8. Graph Neural Networks

Notes

• Foundations: neural networks from scratch
• Design choices: modeling, architecture, and data
• Sequence models:
  • recurrent neural networks
  • decoder-only transformers
  • encoder-decoder transformers
• Vision and generative models:
  • convolutional neural networks
  • diffusion models
• Structured domains: graph neural networks

Executable Companions

• Neural networks from scratch includes an MNIST training experiment and a separate manual-gradient check against autodiff.
• Recurrent neural networks includes a character-level WikiText-2 experiment.
• Decoder-only transformers includes a small causal language-model training and attention experiment.

Sources

• Rumelhart, Hinton, and Williams (1986), Learning Representations by Back-Propagating Errors
• Vaswani et al. (2017), Attention Is All You Need
• LeCun et al. (1998), Gradient-Based Learning Applied to Document Recognition
• Ho, Jain, and Abbeel (2020), Denoising Diffusion Probabilistic Models