Exploratory Data Adventures!

---

     

Open Source Learning and the Democratization of AI

The greatest aspect of Artificial Intelligence today is its open source nature. Very rarely in history have key technologies, especially those as powerful as these models developed today, have been made completely available to the public. Without the incredible online resources made by all of the contributing scientists and engineers, I never would have learned as much as I know today. This repository acts as a documentation of my own exploration as well as an attempt to teach everything I learn to others. I will try my best to cite everyone and everything I reference that helped me learn everything as well!

The main limitation for researchers in this field today is we normally don't have buckets of GPU's just sitting around to train with! Every example I do will be a proof of concept, but to the best of my ability I will attempt to reproduce any model that is feasible!

Contributions

I am typically more wrong than I am right! If you find any errors in my work, that means there is an error in my knowledge. Please let me know as I want this to be as accurate as possible, but also, I want to learn as much as I can! If you want to contribute anything yourself, just submit a PR and I will review it!

Data Prep

We will be using a couple of datasets in our Deep Learning Adventures!!

• Cats Vs Dogs
• IMBD Movie Reviews
• MNIST Handwritten Digits
• Harry Potter Corups

Ensure you have a /data folder in your root directory of the git repo and run the following to install all datasets

bash download_data.sh 

Extra Datasets

There are a few other datasets that we will use but are inconsistent to automatically download and are used in the more advanced architectures! Just download them from the link and save them in the /data folder! These datasets may also be too large to train in Google Drive so keep that in mind!

• CelebA
• MS-COCO
• ADE20K

Foundations

• Intro to PyTorch: Exploring the Mechanics    

• PyTorch Datasets and DataLoaders    

• Leveraging Pre-Trained Models for Transfer Learning    

• Intro to Vision: Digging into Convolutions & AlexNet    

• The Curse of Backprop: Residual Connections & Implementing ResNet  

• Digging into the Sequences: Classification with Recurrence  

• Lets Write a Story: Recurrence for Text Generation    

• Distributed Training with Huggingface 🤗 Accelerate  

Neural Networks from Scratch

• ManualGrad: Simple Neural Network Implementation
• MyTorch: Building a Simple AutoGrad Engine

Computer Vision

• UNet for Image Segmentation  
• Moving from Convolutions: Vision Transformer  
• Masked Image Modeling with Masked Autoencoders
• JEPA

Natural Language Processing

• Causal Language Modeling: GPT 

• Masked Language Modeling: RoBERTa    

• Building Vision/Language Representations: CLIP

• Attention is All You Need for Language Translation   

• Llama 4   

• Language Diffusion Models   

Audio Processing

• Intro to Audio Processing in PyTorch   
• Audio Processing From Scratch
• Connectionist Temporal Classification Loss 
• Intro to Automatic Speech Recognition    
• Quantized Audio Pre-Training: Wav2Vec2 
• Intro to TTS: Tacotron2 
• Neural Vocoder: HifiGAN 
• EnCodec 
• WaveNet

Generative AI

• AutoEncoders

  • Intro to AutoEncoders   
  • Variational AutoEncoders   
  • Vector-Quantized Variational Autoencoders   

• Quantization Methods

  • Gumbel Softmax Vector Quantization   
  • Residual Vector Quantization   
  • Lookup-Free Quantization

• Generative Adversarial Networks

  • Intro to Generative Adversarial Networks   
  • Deep Convolutional GAN   
  • Wasserstein GAN 
  • SuperResolution with SRGAN
  • Image2Image Translation with CycleGAN

• Diffusion

  • Intro to Diffusion    
  • Text-Conditional Diffusion with Classifier Free Guidance
  • Latent-Space Diffusion
  • DDIM
  • Diffusion Transformers

• Autoregressive Generation

  • PixelRNN
  • PixelCNN 

Dive into Transformers

• Attention Mechanisms

  • Attention is All You Need    
  • Windowed Attention
  • LinFormer
  • ConFormer

• Positional Embeddings

  • Sinusoidal Encodings
  • Rotary Positional Encoding
  • ALiBi

• Feed Forward

  • Mixture of Experts

• Normalization

  • Layer Norm
  • RMS Norm

• Inference Optimization

  • KV Cache
  • Multi-Head Latent Attention

Reinforcement Learning

• Model-Based Learning

  • Policy Iteration   
  • Value Iteration   

• Model-Free Learning

  • Monte Carlo   
  • SARSA   
  • Q Learning   
  • TD(N)   
  • TD(Lambda) + Eligibility Traces   

• Intro to Deep Reinforcement Learning

  A. Value Based Methods

  • Deep Q Learning   
  • Double Deep Q Learning   
  • Dueling Deep Q Learning   
  • Prioritized Experience Replay
  • PER With SumTree

  B. Policy and Critic Based Methods

  • Policy Networks (REINFORCE)   
  • REINFORCE with Baseline
  • Generalized Advantage Estimation (GAE)
  • Vanilla Actor-Critic

• Advanced Deep Reinforcement Learning

  • Synchronous Actor-Critic (A2C)
  • Trust Region Policy Optimzation (TRPO)
  • Proximal Policy Optimization (PPO)
  • Deep Deterministic Policy Gradient (DDPG)
  • Twin Delayed DDPG (TD3)
  • Soft Actor Critic (SAC)

Tools

• Gradient Checkpointing
• Low Rank Adaptation     
• Knowledge Distillation
• Quantized Low Rank Adaptation
• TensorRT
• DeepSpeed
• Triton