Complete AI/ML Study Plan
Anas Hussain (a1n4a) — Frontier Mastery Programme
Based on : All documents in the AI_ML libraryDuration : 24 MonthsGoal : Frontier-level research competence
Your Document Library (Reading Order)
Priority
Document
Focus
When
1
AI_ML_Mastery_Roadmap
Strategy & habits
Week 1
2
a1n4a_deep_understanding
Conceptual depth
Months 1-4
3
a1n4a2
Zero-to-researcher arc
Months 1-6
4
a1n4a_complete_textbook
Full technical reference
Ongoing
5
research_summary_phases_1_3
Cutting-edge theory
Months 5-8
6
phases_4_5_6_research_summary
RLHF / LLMs / MoE
Months 9-14
7
phases_7_8_9_research_summary
TCS / Neuro / Frontier
Months 15-20
8
AI_ML_Frontier_Research
Pure research reference
Months 6-24
Phase 1: Foundations (Months 1-4)
Week 1-2: Orientation & Audit
[ ] Read AI_ML_Mastery_Roadmap cover to cover
[ ] Complete the self-assessment checklist — mark every gap
[ ] Set up your Known Unknown Map file
[ ] Install: Python, PyTorch, Jupyter, VS Code
Week 3-6: Mathematical Core
Primary : a1n4a_deep_understanding Ch 2-4Reference : a1n4a_complete_textbook Part 0
Week 7-10: Classical ML
Primary : a1n4a_deep_understanding Ch 5-8 + a1n4a2 Sec 1-4Reference : a1n4a_complete_textbook Ch 1-8
[ ] Linear Regression — implement gradient descent + closed form
[ ] Logistic Regression — implement + sigmoid derivation
[ ] SVMs — understand margin maximization, kernel trick
[ ] Decision Trees + Random Forests — implement CART
[ ] Theory of Generalization — VC dimension, PAC learning
Weekly Deliverable : Implement each algorithm from scratch. No sklearn.
Week 11-16: Statistical Learning Theory
Primary : AI_ML_Frontier_Research Sec 1Reference : a1n4a_complete_textbook Ch 7
[ ] PAC Learning framework — prove the main theorem
[ ] VC dimension — prove shattering for linear classifiers
[ ] Rademacher Complexity — derive the generalization bound
[ ] Double Descent — reproduce Belkin et al. 2019 experiment
[ ] NTK theory — understand the infinite-width limit
Phase 1 Gate Test : Implement backpropagation in pure NumPy for a 3-layer network. Verify gradients with finite differences.
Phase 2: Deep Architectures (Months 5-8)
Month 5: Neural Networks Deeply
Input
Hidden 1
Hidden 2
Output
x₁
x₂
x₃
x₄
ŷ₁
ŷ₂
4 neurons
5 neurons
4 neurons
2 neurons
→ Forward Pass →
Fig. 1 — Feedforward Neural Network: 4-input, two hidden layers, 2-output. Each circle = neuron; each edge = learnable weight.
Primary : a1n4a_deep_understanding Ch 9+Reference : a1n4a_complete_textbook Ch 9-14
[ ] Implement a full neural network class in NumPy (no PyTorch)
[ ] Derive and implement backpropagation — match analytical and numerical gradients
[ ] Implement BatchNorm, LayerNorm, Dropout
[ ] Implement SGD, Adam, RMSProp — compare convergence
[ ] Study activation functions — ReLU, GELU, Swish — theoretical comparison
Transformer Encoder Block
Input Embeddings
+ Positional Encoding
Multi-Head Attention
Q
K
V
Attention(Q,K,V) = softmax(QKᵀ/√dₖ)V
Add & LayerNorm
Feed-Forward Network
Linear → ReLU → Linear
dim: 512 → 2048 → 512
Add & LayerNorm
Output (to next block)
Residual
×N
stacked
blocks
Fig. 2 — Transformer Encoder Block: Multi-Head Attention + residual connections + Layer Normalization + Feed-Forward Network.
Primary : a1n4a_complete_textbook Ch 16Reference : research_summary_phases_1_3 Sec Phase 2
[ ] Read "Attention Is All You Need" — annotate every equation
[ ] Implement Multi-Head Attention from scratch (NumPy)
[ ] Implement Positional Encoding: absolute, RoPE, ALiBi
[ ] Implement Flash Attention — understand IO complexity
[ ] Build a complete mini-GPT (< 200 lines)
[ ] Mechanistic analysis: find induction heads in your model
Month 7: CNN, GNN, SSM
Primary : a1n4a_complete_textbook Ch 11, 20, 21
[ ] Implement CNN — derive convolution as correlation
[ ] Group equivariance — understand why CNNs work on images
[ ] Implement a basic GNN — message passing
[ ] Weisfeiler-Lehman isomorphism test — prove GNN limitation
[ ] State Space Models (S4/Mamba) — implement the recurrence
Month 8: Generative Models
Primary : a1n4a_complete_textbook Ch 22-27Reference : research_summary_phases_1_3 Sec Phase 3
[ ] VAE — derive ELBO from KL divergence; implement
[ ] GAN — derive minimax; implement; observe mode collapse
[ ] DDPM — derive forward/reverse SDE; implement full training
[ ] Flow Matching — understand the vector field formulation
[ ] Normalizing Flows — implement Real-NVP
Phase 2 Gate Test : Implement DDPM from scratch. Train on MNIST. Generate samples.
Phase 3: Learning Paradigms (Months 9-12)
Month 9: Self-Supervised & Contrastive Learning
Contrastive Self-Supervised Learning (SimCLR / CLIP)
Image x
Augment 1
crop+flip+color
Augment 2
grayscale+blur
Encoder f
z₁
Encoder f
z₂
Project +
Compare
InfoNCE Loss
Pull z₁,z₂ together
Push negatives apart
No labels needed — augmented views of same image = positive pairs; different images = negatives.
Fig. 15 — Contrastive SSL: two augmented views of the same image form a positive pair. InfoNCE loss pulls them together in representation space, pushes negatives apart.
Primary : a1n4a_complete_textbook + phases_4_5_6_research_summary
[ ] Derive InfoNCE loss from mutual information
[ ] Implement SimCLR — train on CIFAR-10 without labels
[ ] Study why BYOL works without negative pairs (theory)
[ ] Implement CLIP at small scale
Month 10: Reinforcement Learning
Reinforcement Learning — Agent-Environment Loop
Agent
Policy π(a|s)
Value V(s)
Environment
State transitions
Reward function R(s,a)
Action aₜ
State sₜ₊₁ + Reward rₜ
Goal: maximise cumulative reward G = Σ γᵗ rₜ
Fig. 10 — RL Agent-Environment loop: agent selects actions via policy π; environment returns next state and reward; agent maximises cumulative discounted reward.
Primary : a1n4a_complete_textbook Ch 30
[ ] Implement Q-Learning on FrozenLake (OpenAI Gym)
[ ] Derive Policy Gradient Theorem — prove REINFORCE
[ ] Implement PPO from scratch — understand clipping
[ ] Implement Actor-Critic (A3C)
[ ] Read Sutton & Barto — full book
Month 11: RLHF & Alignment
RLHF — Reinforcement Learning from Human Feedback
Step 1
Supervised
Fine-Tuning
(SFT on demos)
Step 2
Reward Model
Training
(human pref pairs)
Step 3
PPO Training
with RM
(policy optimisation)
Aligned
LLM
Human Annotators
rank model outputs
Alternative: DPO
Direct Pref. Optimisation
(skips explicit RM)
Fig. 9 — RLHF Pipeline: (1) SFT on demonstrations, (2) train Reward Model from human preferences, (3) PPO to optimise policy against RM.
Primary : phases_4_5_6_research_summary Sec 4.1Reference : a1n4a_complete_textbook Ch 31-32
[ ] Read InstructGPT paper — annotate every contribution
[ ] Derive DPO from RLHF objectives — mathematical proof
[ ] Implement toy RLHF pipeline (reward model + PPO)
[ ] Study Constitutional AI method
Month 12: Advanced Paradigms
[ ] MAML — derive bi-level optimization; implement on Omniglot
[ ] EWC — derive Fisher Information regularizer
[ ] IRM — derive invariant risk minimization objective
[ ] LoRA — understand low-rank decomposition
Phase 3 Gate Test : Train a small language model with RLHF on a custom reward function.
Phase 4: Foundation Models (Months 13-16)
Month 13: LLM Internals
Primary : a1n4a_complete_textbook Ch 28Reference : phases_4_5_6_research_summary Sec 5
[ ] Study nanoGPT source line by line
[ ] Rederive Chinchilla scaling laws
[ ] Implement BPE tokenizer from scratch
[ ] Analyze data quality research: FineWeb, Dolma
Month 14: Reasoning & Test-Time Compute
Primary : phases_4_5_6_research_summary Sec 5.2
[ ] Read DeepSeek-R1 — understand RL for reasoning
[ ] Study Chain-of-Thought: why does it improve accuracy?
[ ] Process Reward Models vs Outcome Reward Models
[ ] Study MCTS applications in language models
Month 15: Multimodal + MoE
Primary : a1n4a_complete_textbook Ch 29, 34
[ ] CLIP implementation at small scale
[ ] Study Flamingo cross-attention architecture
[ ] MoE sparse routing — derive load balancing loss
[ ] Implement routing layer
Month 16: Mechanistic Interpretability
Primary : AI_ML_Frontier_Research Sec 6Reference : phases_7_8_9_research_summary
[ ] Read Anthropic "Toy Models of Superposition"
[ ] Train a Sparse Autoencoder (SAE) on GPT-2 activations
[ ] Find induction heads in your mini-GPT
[ ] Study polysemanticity in MLP neurons
Phase 4 Gate Test : Write a 3-page technical critique of a recent NeurIPS 2024 paper.
Phase 5: Safety & Frontier (Months 17-24)
Month 17-18: AI Safety Theory
Primary : AI_ML_Frontier_Research Sec 6
[ ] Read Stuart Russell "Human Compatible" — full book
[ ] Study mesa-optimization and inner alignment
[ ] Study deceptive alignment — formal models
[ ] Certified robustness — implement randomized smoothing
Month 19-20: TCS & Theoretical Connections
Primary : phases_7_8_9_research_summary
[ ] Computational complexity meets ML — lower bound proofs
[ ] Cryptographic hardness in ML security
[ ] Quantum ML — critically evaluate claimed advantages
[ ] Circuit complexity and neural network expressivity
Month 21-24: Your Frontier
Pick 2-3 from the frontier areas and go deep:
[ ] Geometric DL — implement SE(3)-equivariant model
[ ] Causal AI — implement causal discovery algorithm
[ ] World Models — implement JEPA-style latent prediction
[ ] Scientific AI — fine-tune ESM2 for protein property prediction
[ ] AI Agents — build a multi-step tool-using agent
Daily Schedule
Time
Activity
07:00-07:45
Read 1 arxiv paper — write 3 bullet summary
07:45-08:00
Implement or derive 1 thing from yesterday's paper
20:00-21:30
Deep work: derive, code, or write
21:30-22:00
Update Known Unknown Map
22:00-22:15
Queue tomorrow's paper (arxiv.org/list/cs.LG/recent)
Weekly Review Template
Every Sunday:
Benchmark Tests
Month
Test
4
Implement backpropagation from scratch. Match finite-difference gradients.
8
Implement DDPM. Train on MNIST. Generate 64 samples.
12
Full RLHF pipeline: SFT + RM + PPO. It must converge.
16
3-page critique of a recent paper. Identify one extendable contribution.
20
State one unsolved problem. Propose an experiment. Write it up.
24
Submit to a workshop (NeurIPS/ICML/ICLR).
Resources by Phase
Phase 1 (Foundations)
Understanding Machine Learning — Shalev-Shwartz & Ben-DavidMathematics for ML — Deisenroth et al. (free PDF)Information Theory, Inference & Learning — MacKay (free PDF)
Phase 2 (Architectures)
Deep Learning — Goodfellow, Bengio, CourvilleAndrej Karpathy's YouTube (Neural Nets: Zero to Hero)
"Attention Is All You Need" — Vaswani et al. 2017
Phase 3-4 (Learning + Foundation Models)
Reinforcement Learning — Sutton & Barto (2nd ed., free PDF)InstructGPT paper (Ouyang et al., 2022)
DeepSeek-R1 technical report (2025)
Phase 5 (Safety + Frontier)
Human Compatible — Stuart RussellAnthropic Research Blog (mechanistic interpretability)
Alignment Forum (alignmentforum.org)
The Measurement: Your Known Unknown Map
Update this file weekly. The goal is to move everything from the bottom to the top.
KNOWS DEEPLY ( can derive + implement + explain without notes ):
- [ ] Backpropagation
- [ ] Self - attention mechanism
- [ ] ELBO derivation for VAE
...
KNOWS SUPERFICIALLY ( can explain but not derive ):
- [ ] NTK theory
- [ ] Wasserstein GAN
...
KNOWS EXISTS ( but cannot explain ):
- [ ] Free Energy Principle
- [ ] Kolmogorov complexity
...
"The goal is not to finish this plan. The goal is to become the kind of person who never stops."
Study Plan generated from AI/ML document library — July 2026 Documents covered: a1n4a_complete_textbook, a1n4a_deep_understanding, a1n4a2, AI_ML_Frontier_Research, AI_ML_Mastery_Roadmap, phases research summaries