Machine Learning from Scratch

Series Information

Program: Machine Learning from Scratch
Institution: Harvard Medical School
Year: 2021
Organizers: Johannes Bill, John Vastola, Binxu Wang

Target Audience: Neuroscience graduate students and postdocs

Series Objective

A student-led tutorial and seminar series designed to teach neuroscience students cutting-edge machine learning models by implementing them from scratch. The hands-on approach ensures deep understanding of both theoretical foundations and practical implementation.

Tutorial Topics

1. Transformers and Large Language Models (LLMs)

Topics Covered:

• Natural Language Processing (NLP) fundamentals
• Attention mechanism deep dive
• Transformer architecture from first principles
• Training language models
• Usage and fine-tuning of pretrained models
• Applications beyond language: vision, audio, image generation

Materials:

• Interactive Jupyter/Colab notebooks
• Step-by-step implementation guides
• Detailed mathematical derivations
• Practical coding exercises

2. Stable Diffusion Models

Topics Covered:

• Principles of diffusion models
• UNet model architecture
• Contextualized word embedding
• Cross-attention mechanisms
• Autoencoder efficiency techniques
• Large-scale training strategies

Key Concepts:

• Forward and reverse diffusion processes
• Noise scheduling
• Conditional generation
• Image synthesis applications

3. Mathematical Foundation of Diffusion Generative Models

Advanced Topics:

• Score function and data distribution gradients
• Reversing forward diffusion process mathematically
• Score function learning techniques
• Neural network score approximation
• Advanced sampling methods (DDPM, DDIM)

Materials:

• Rigorous mathematical derivations
• Implementation from theoretical principles
• Performance optimization techniques
• Comparison of different sampling strategies