Machine and Deep Learning Study Roadmap

Due to Git LFS problems, all 67 commits that included the repo's history had to be deleted :(

Overview

This roadmap includes resources, courses and their projects, and notes that I am following to build a strong foundation in machine and deep learning.

Contents

• Prerequisites: Quick revision on NumPy, Pandas, and Matplotlib.
• MIT 6.S191: Introduction to Deep Learning.
• PyTorch Course
• Projects: Contains most course projects; larger projects have their own repos linked below.
• LLM from Scratch Course
• ITI Summer Training: Machine Learning (supervised and unsupervised).
• Graduation Project Notes: Research notes from my Bachelor’s degree project on Parkinson’s disease (not the actual project, but concepts explored and ML-related notes for future reference).

Resources & Links

Roadmap Video:

• How to Learn Machine Learning in 2024 (7-step roadmap) (contains links to many resources)

Courses:

• MIT 6.S191: Introduction to Deep Learning (highly recommended)
• Full PyTorch Course
  • GitHub Repo
  • Book Page
• LLMs from Scratch
• Stanford Introduction to Machine Learning (free on Udacity)
• Statistical Learning with Python

Online Tutorials & Blogs:

• All Machine Learning Algorithms Explained in 17 Min
• 15 Machine Learning Lessons I Wish I Knew Earlier
• Cracking Machine Learning Interview: Repository to prepare for ML interviews, covering common questions asked by top companies.
• Great ML/DL YT channel (StatQuest with Josh Starmer)

Useful Websites:

• Browse State-of-the-Art
• Hugging Face Models
• timm (PyTorch Image Models) Library

Projects & Practice

• Older projects in this repo are mostly drafts and unpolished experiments.
• More recent projects were developed in their own dedicated repositories (including experiments and failed attempts), and are no longer included in this repo.
• For complete projects, please check the linked repositories.

Main Projects

• Project Mozart: Custom Transformer-based music generation.
• Brain Tumor Segmentation: Deep learning model for semantic segmentation of brain tumors from MRI scans.
• German Tutor: AI-powered German language learning assistant.
• CelluScan: Automated blood cell classification with Vision Transformer.
• Real-Time Number Detector Web App: Trained on MNIST + custom data.

PyTorch Course Tasks

• Fashion MNIST Model (with confusion matrix, Chapter 3)
• Multi-class Evaluation Model (Non-linear, Chapter 2)
• Linear Regression Model (Chapter 1)
• Final Project:
  • Food MINI Model: Experiment tracking and model deployment.

---

Machine Learning & Deep Learning Mastery Checklist

1. Mathematical Foundations (from college)

• Linear Algebra (vectors, matrices, eigenvalues, SVD)
• Calculus (gradients, Jacobians, chain rule)
• Probability & Statistics (distributions, expectation, Bayes)
• Optimization (gradient descent, Adam, RMSprop, etc.) (next term)

2. Classical Machine Learning

• Linear Regression
• Logistic Regression
• Decision Trees / Random Forests
• Boosting and Bagging
• Support Vector Machines (SVM)
• K-Nearest Neighbors (KNN)
• Naive Bayes (Multinomial, Gaussian)
• Gradient Boosting / XGBoost
• K-Means & other clustering methods (Hierarchical clustering, DBSCAN)
• Gaussian Mixture Models (GMM)
• Dimensionality Reduction (PCA)
• Bias–Variance Tradeoff
• Cross-validation
• Regularization (Ridge (L2) and Lasso (L1))
• Evaluation Metrics (Accuracy, Precision, Recall, F1, ROC, AUC)
• Feature Engineering / Normalization

3. Core Deep Learning

• Perceptron, Feedforward Networks
• Activation Functions (ReLU, Sigmoid, Tanh, GELU, etc.)
• Backpropagation
• Weight Initialization (Xavier, Kaiming)
• Dropout / BatchNorm / LayerNorm
• Loss Functions (MSE, Cross-Entropy, BCE, etc.)
• Learning Rate Schedules
• Early Stopping
• Gradient Clipping
• Data Augmentation
• Mixed Precision Training

4. Convolutional Neural Networks (CNNs)

• Convolution / Pooling / Padding
• LeNet
• AlexNet
• VGG
• ResNet
• EfficientNet
• Transfer Learning
• Object Detection (YOLO)
• Segmentation (U-Net, Mask R-CNN)

5. Recurrent and Sequential Models

• RNNs
• LSTMs
• GRUs
• Sequence-to-Sequence Models
• Attention Mechanisms
• Encoder–Decoder Architectures

6. Transformers and Attention Models

• Self-Attention & Multi-Head Attention
• Transformer Architecture (Encoder, Decoder)
• Vision Transformer (ViT)
• BERT & GPT-style models (encoder vs decoder-only)
• Positional Encoding
• Fine-tuning large models
• Top-k, Top-p (nucleus) sampling

7. Generative Models

• Autoencoders (AE)
• Variational Autoencoders (VAE)
• Generative Adversarial Networks (GANs)
• Diffusion Models (DDPM, Stable Diffusion, ControlNet)
• Flow-based Models (RealNVP, Glow)

8. Self-Supervised / Representation Learning

• Contrastive Learning (SimCLR, MoCo)
• Masked Autoencoders (MAE, BEiT)
• DINO / CLIP (cross-modal self-supervision)

9. Reinforcement Learning (RL)

• Markov Decision Processes (MDPs)
• Policy Evaluation / Improvement
• Value Iteration / Policy Iteration
• Temporal Difference Learning
• Monte Carlo Methods
• Deep Q-Network (DQN)
• Double / Dueling DQN
• Policy Gradient Methods
• Actor–Critic
• PPO (Proximal Policy Optimization)
• DDPG / TD3 / SAC (continuous control)
• Model-Based RL (Dreamer, MuZero)
• Multi-Agent RL (MADDPG, QMIX)
• Offline RL / Imitation Learning
• RLHF (Reinforcement Learning from Human Feedback)

10. AI Agents & Autonomous Systems

• ReAct (Reason + Act) architecture
• AutoGPT / BabyAGI
• Retrieval-Augmented Generation (RAG)
• LangChain or LlamaIndex for tool-using agents
• Vector DB Memory (FAISS, Chroma, Milvus)
• Planning + Reflection loops
• Integrating RL / feedback into agents
• Multi-agent coordination and communication

11. Deployment & MLOps

• Hugging Face deployment
• TorchScript / ONNX model export
• FastAPI or Flask inference servers
• Dockerization for reproducible deployment
• GPU inference optimization (mixed precision, batching)
• Model quantization / pruning
• Distributed training (PyTorch DDP / DeepSpeed)
• CI/CD for models (GitHub Actions + versioning)
• Monitoring / logging in production (W&B, Mlflow, TensorBoard)

12. Practical / Projects

• Build and train models from scratch
• Reproduce architectures (ViT, Transformer, etc.)
• Train a symbolic music generation Transformer
• Visualize training metrics (TensorBoard, W&B)
• Use pretrained models (Hugging Face)
• Handle datasets and dataloaders efficiently
• Implement an RL environment and agent
• Build a tool-using AI agent with memory