Phase Ladder

Study Plan

This is the academy's single map. Follow the six core phases in order if you are learning AI for the first time. Use them as a repair map if you are preparing for IOAI and entering at your weakest layer. Each core phase has an exit gate — you should be able to pass it cold before moving on. The elective modules after Phase 6 are chosen by need and have no shared gate.

First Steps IOAI Surface Tracks Decision Clinics

New To AI

Start At Phase 1

Do the phases in order. Do not skip. Each phase produces one workflow habit the next phase assumes.

Preparing For IOAI

Enter At Your Weakest Phase

Read the exit gate for each phase. Enter at the first one you cannot pass cold. Do not restart Phase 1 out of habit.

Already Advanced

Maintenance Mode

When all six core exit gates are easy, choose a specialized elective or move to Beyond The Academy. The core is a floor, not a ceiling.

Before Phase 1

You should already be able to:

• create and use a Python virtual environment
• run one script from the terminal
• read a short traceback
• tell the difference between a topic page, an example, and a track

If any of that is shaky, start with First Steps and then Getting Started.

The run commands below assume macOS or Linux and the academy repository root. On Windows PowerShell, replace .venv/bin/python with .venv\Scripts\python.exe; keep the examples/... path relative to the repository root.

First 90 Minutes

If you want a single short opening before committing to a full phase:

1. read Getting Started
2. run one small example from Examples
3. inspect one artifact carefully — do not just confirm the script ran
4. read the exit gate of Phase 1 below
5. if you cannot pass the gate cold, start at Phase 1; otherwise start at the first phase you cannot pass

If you are preparing for IOAI, also open IOAI Competition Surface once to understand how problems are delivered, then come back here.

Phase 1 — Tooling And Inspection

Goal. Make arrays, tables, grouped summaries, and plotting feel mechanical enough that the first model does not break your attention.

Do all of this:

1. What Is a Model?
2. Probability and Linear Algebra Refresher
3. Array Shapes and Axis Operations
4. Table Inspection
5. Data Cleaning and Preprocessing
6. Grouped Summaries and Slice Checks
7. Plotting for Model Debugging
8. Feature Matrix Construction
9. run .venv/bin/python examples/numpy-shape-sanity/shape_sanity.py
10. run .venv/bin/python examples/course-support-inspection/quick_inspection.py
11. Track: Python, NumPy, Pandas, Visualization

Exit gate — move on when you can:

• explain row versus column operations without thinking
• inspect missing values and label balance from a cold table in under two minutes
• build a simple feature matrix without losing row alignment
• make one plot that changes what you inspect next
• explain what a model learns and why evaluation needs unseen examples

Phase 2 — Honest Validation

Goal. Make split discipline, baselines, cross-validation, tuning, calibration, and leakage checks feel non-negotiable. Nothing else in the academy matters until this phase is automatic.

Do all of this:

1. Honest Splits and Baselines
2. Leakage Patterns
3. Cross-Validation
4. Hyperparameter Tuning
5. run .venv/bin/python examples/validation-baseline-comparison/baseline_comparison.py
6. run .venv/bin/python examples/classical-ml-recipes/leakage_patterns_demo.py
7. Clinic: Public/Private Restraint — do it cold, write the note, then check the reveal
8. Clinic: Leakage Or Signal?
9. Validation, Leakage, and Model Choice — answer every question before looking at solutions
10. Track: scikit-learn Validation and Tuning — full run with artifacts

Exit gate — move on when you can:

• explain why the test set must stay untouched after model selection
• name three leakage patterns without looking them up
• defend a model choice using paired fold deltas, not just the mean
• pass Sheet A: Validation and Leakage Sprint cold

Phase 3 — Model Diagnostics And Classical Choice

Goal. Diagnose model behavior and make justified classical model choices. Add geometry, clustering, and representation judgment without losing evaluation discipline.

Do all of this:

1. Linear Regression
2. Regression Metrics and Diagnostics
3. Logistic Regression
4. K-Nearest Neighbors
5. Decision Trees
6. SVM Margins and Kernels
7. Clustering and Low-Dimensional Views
8. K-Means
9. PCA
10. Dimensionality Reduction
11. Advanced Clustering and Dimensionality Reduction
12. Feature Selection
13. Learning Curves and Bias-Variance
14. Calibration and Thresholds
15. Ensemble Methods
16. Evaluation Metrics Deep Dive
17. Clinic: Overfit Or Underfit?
18. Clinic: Ensemble Temptation
19. Clinic: Threshold Under Asymmetric Cost
20. run .venv/bin/python examples/classical-ml-recipes/calibration_threshold_demo.py
21. SVM, Kernels, and Learning Theory
22. Track: SVM and Advanced Clustering

Exit gate — move on when you can:

• diagnose overfit vs. underfit from a training curve in under a minute
• choose a threshold when false negative cost is 100× false positive cost
• reject an ensemble that wins by a marginal amount with high operational cost
• justify linear versus nonlinear boundaries on a given dataset
• explain when cross-validation is selection evidence vs. final evidence

Phase 4 — Deep Learning Control

Goal. Make training loops, checkpoints, optimizer choices, and overfitting checks feel controlled instead of mysterious. Know when the training loop itself is the bug.

Do all of this:

1. Loss and Gradients Intuition
2. MLP Training Baseline
3. Backpropagation
4. Activation Functions
5. PyTorch Training Loops
6. Optimizers and Regularization
7. Batch Normalization and Initialization
8. Learning Rate Schedulers
9. Debugging Deep Learning
10. Clinic: Checkpoint Roulette
11. run .venv/bin/python examples/deep-learning-recipes/pytorch_training_loop_demo.py
12. run .venv/bin/python examples/deep-learning-recipes/optimizer_regularization_demo.py
13. Lab: labs/cnn-from-scratch — receptive fields, translation equivariance, and overfit-one-batch checks
14. Deep Learning and Checkpoints
15. Track: PyTorch Training Recipes — full run with checkpoint selection

Exit gate — move on when you can:

• describe the difference between training and evaluation mode without hesitating
• select the right checkpoint from a 15-epoch log
• explain why training loss is irrelevant for checkpoint selection
• name the first three things to check when a training run diverges
• pass Sheet C: Training and Checkpoint Sprint cold

Phase 5 — Transfer And Representation

Goal. Decide when to freeze, probe, fine-tune, or switch representation families instead of retraining everything from scratch. Compare backbones honestly.

Do all of this:

1. Transfer and Fine-Tuning
2. Convolutional Neural Networks
3. Attention and Transformers
4. Vision and Text Encoders
5. Self-Supervised and Representation Learning
6. Metric Learning and Retrieval
7. Semi-Supervised Learning
8. Clinic: Freeze Or Fine-Tune?
9. run .venv/bin/python examples/deep-learning-recipes/transfer_finetuning_demo.py
10. Lab: labs/self-supervised-representation — SimCLR-style pretraining, linear probe at multiple label budgets
11. Lab: labs/clip-multi-modal — symmetric InfoNCE, image↔text retrieval asymmetry
12. Lab: labs/metric-learning-retrieval — triplet / batch-hard / prototype classifiers, Recall@k on unseen identities
13. Lab: labs/semi-supervised-learning — pseudo-labeling vs. label spreading on a 20-labeled + 2000-unlabeled pool
14. Track: Synthetic Transfer and Fine-Tuning
15. Track: Representation Reuse and Embedding Transfer

Exit gate — move on when you can:

• justify frozen features vs. partial fine-tuning vs. full fine-tuning given a data budget
• compare two pretrained backbones on the same downstream task with honest evaluation
• explain when fine-tuning destroys more than it gains
• pass Sheet B: Representation and Geometry Sprint cold

Phase 6 — Decisions Under Constraint

Goal. Practice model choice, operating points, and workflow discipline under time, budget, or leaderboard pressure. This is where IOAI readiness is built.

Do all of this:

1. Baseline-First Task Solving
2. Mock Tasks and Timed Workflows
3. Imbalanced Metrics and Review Budgets
4. Selective Prediction and Review Budgets
5. Experiments and Ablations
6. Reliability Slices
7. Clinic: Review Budget Freeze
8. run .venv/bin/python examples/mock-task-recipes/baseline_first_demo.py
9. run .venv/bin/python examples/mock-task-recipes/review_budget_demo.py
10. Track: Vision and Audio Workflows
11. Track: Mock Tasks and Timed Workflows — full run under self-imposed time constraint
12. Track: Imbalanced Triage and Review Budgets
13. Timed Checkpoint Sheets — all three sheets, timed, no notes

Exit gate — move on when you can:

• produce a baseline, one iteration, and a stop decision within a fixed time window
• choose a review budget policy that survives a queue constraint
• defend a stop call when the visible score still has room to improve
• reject a public leaderboard jump that local validation does not support

If you are preparing for IOAI, pair this phase with IOAI Competition Surface for problem-reading, sandbox, and submission practice.

Elective Modules — Specialized Modalities

These modules sit outside the six-phase core and do not share an exit gate. Choose the section that matches a task you need to solve; do not read them linearly.

Goal. Extend the core workflow into specialized modalities and advanced techniques without losing split, baseline, inspection, or stop discipline.

Vision And Detection

• Vision Augmentation and Shift Robustness
• Object Detection Basics
• Detection and Segmentation
• Density-Based Counting
• Saliency and Explainability
• Lab: labs/density-map-counting — scalar-count vs. detect-then-count vs. density regression, σ sweep
• Lab: labs/saliency-and-explainability — Grad-CAM, Integrated Gradients, pointing-game scoring, rectangle-from-heatmap
• Track: Detection and Segmentation Workflows

Text And Language

• Text Representations and Order
• Tokenization Mechanics
• Language Modeling Fundamentals
• Text Generation and Language Models
• Prompting and Tool Use
• Retrieval-Augmented Generation
• Encoder-Decoder and Translation
• Recurrent Networks and Sequences
• Clinic: Prompt Vs Retrieval Vs Fine-Tune — do it cold before the RAG lab
• Clinic: Published Result Trust — before anchoring your work on a benchmark number
• Black-Box LLM Optimization
• Lab: labs/encoder-decoder-translation — GRU+attention seq2seq, teacher-forced vs free-running loss, greedy vs beam decoding
• Track: Text Workflows Beyond Classification
• Track: RAG And Prompting Workflows — two-loop evaluation with retrieval and rerank attribution

Generative Models

• Autoencoders and VAEs
• Generative Adversarial Networks
• Diffusion Models
• Steering Frozen Generative Models
• Lab: labs/autoencoder-and-vae — reconstruction vs. generation, β trade-off, posterior collapse
• Lab: labs/gan-workflow — vanilla/LSGAN/WGAN-GP, mode coverage on a 4-mode mixture
• Lab: labs/diffusion-models — linear vs cosine schedules, DDPM vs DDIM sampling step budgets
• Lab: labs/generative-steering-and-search — textual inversion to a held-out class, embedding-direction transfer, UCB/beam search against a stochastic scorer, reward-hacking demo

Audio And Speech

• Audio Windows and Spectral Features
• Audio Models
• Track: Speech and Audio Encoders

Multi-Modal And Structured

• Multi-Modal Fusion
• Sequential Splits and Lag Features
• Tabular Feature Engineering

Optimization And Efficiency

• Mixed Precision Training
• PyTorch Optimization Recipes
• PEFT and LoRA
• Mixture-of-Experts and Scaling
• Quantization, Distillation, and Deployment
• Data Augmentation
• Track: Optimization, Regularization, and PEFT

Graphs And Sequential Decisions

• Graph Neural Networks
• Reinforcement Learning Foundations

Post-Model Algorithms And Theory

• Track: Structured Post-Model Algorithms
• Track: Problem Adaptation and Post-Processing
• Track: Theory Checkpoint Sheets
• Unsupervised Learning and Representation

When all six core phases are boring and the electives you need no longer stretch you, move to Beyond The Academy.

Repair Map

If you already know your weakness, skip the phase ladder and go directly:

Weakness	Go here
Splits and leakage are shaky	Phase 2 from the top
Cannot diagnose overfit vs. underfit	Clinic: Overfit Or Underfit? then Phase 3
Training loops feel fragile	Phase 4 from the top
Transfer decisions are guesses	Clinic: Freeze Or Fine-Tune? then Phase 5
Decisions collapse under time pressure	Phase 6, start with the mock task track
Threshold/operating-point choices are weak	Clinic: Threshold Under Asymmetric Cost then Phase 6
Ensemble/complexity choices are weak	Clinic: Ensemble Temptation then Phase 3
Need a specific domain (vision, text, audio)	Elective Modules, pick the matching section
IOAI problem format is unfamiliar	IOAI Competition Surface, then Phase 6