AI · Overview & Knowledge Map¶
Broad coverage, focused study
Distinguish foundations, methods, tasks and systems, then choose an implementation track. P0 is shared foundation, P1 major-method literacy, P2 specialization. A comprehensive map is not a requirement to complete every topic.
Structure¶
flowchart TD
A[Learning foundations and classical methods] --> B[Representations and architectures]
B --> C[Generation and perception]
B --> D[Search planning and decision learning]
C --> E[Foundation models and multimodal applications]
D --> F[Robotics and embodied intelligence]
B --> G[Time series and scientific ML]
H[Data and evaluation] -.-> A
H -.-> C
H -.-> D
I[Efficient training and deployment] -.-> B
I -.-> E
I -.-> F| Module | Coverage | First pass |
|---|---|---|
| Foundations and classical ML | Probability/optimization, generalization, data, linear/kernel/tree/ensemble methods, clustering and reduction | Diagnose gaps and build a sound baseline |
| Representation and paradigms | Embeddings, transfer, self-supervision, contrast/masking/teachers, continual/active/federated learning | Explain supervision and compare feature use |
| Architectures | MLP, CNN, RNN, Transformer, SSM, MoE and efficient attention | Structural trade-offs and one implementation |
| Graphs and geometry | Message passing, graph Transformers, symmetry/equivariance, points and relations | Purpose first; GCN after selecting graphs |
| Generation | Autoregression, VAE, GAN, normalizing flows, diffusion, flow matching and discrete generation | Objectives and sampling; optional 2D experiments |
| Vision, multimodality and space | Classification, detection, segmentation, video, audio/language alignment, depth/points, NeRF/3DGS | Tasks and metrics before 3D depth |
| Search and decision learning | Search, constraints, knowledge, RL, imitation/offline learning, world models and skills | Prediction vs. action; reuse robot practice |
| Time series and scientific AI | Forecasts/anomalies, transfer, PINNs, neural operators and inverse problems | Reuse signals/systems experience and classical baselines |
| Efficient systems | Reproducibility, profiling, quantization/distillation/pruning, parallelism, runtimes and deployment | Measure resources on an existing model |
| Evaluation and trustworthy AI | Data quality, calibration, uncertainty, OOD, explanations, privacy/safety and bias | Use throughout every project |
AI covers transferable methods; existing sections retain LLM lifecycle and robot-loop detail rather than duplicate courses.
Resource selection¶
Continue Lee/Karpathy as existing backbones. Use CS231n for vision and representations and MIT 6.S184 for generation. Enter graph/RL courses only after selecting those directions; use official walkthroughs for small experiments.
Each topic normally has one primary entry and one necessary practical/reference entry. Topics follow a consistent cognitive loop: Knowledge Map \(\to\) Core Mechanistic Intuition \(\to\) Selected Entries \(\to\) Controlled Guided Experiment \(\to\) Check-Yourself Reflection. Intuition anchors the concepts, experiments verify the code, and self-checks test deep understanding. Sources, scope and difficulty checked 2026-10-18; names and popularity counts alone are not quality evidence.
The original checklist remains a historical record. Track later ability through explain/reproduce/independently-improve evidence.