Learning paths

Build a solid mathematical and conceptual grounding. Covers linear models, error decomposition, and essential optimization.

Trace the evolution up to modern landmarks. Covers CNNs, initialization, and Memory Networks.

Construct a journey from basic decision processes to aligning complex models.

Teach how ML systems "see" and map vision features to text.

Explore the bleeding edge of deployment, pruning, and safety mechanisms for advanced users.

Rare but high-value theoretical foundations. Move from PAC learning and VC dimension to modern insights like double descent and neural collapse.

The engineering behind large models. Scaling laws, tokenization, KV cache optimization, and high-performance serving.

Inducing System 2 thinking. From in-context learning to chain-of-thought and learned tool-use architectures.

Ensuring models behave predictably. RLHF, DPO, reward modeling, and the frontiers of mechanistic interpretability.

Modern generative engines. Score matching, SDEs, and the transition to latent diffusion and video generation.

Beyond vanilla attention. Explore Mixture-of-Experts (MoE), linear attention, and state-space models like Mamba.

Inference and training efficiency. Quantization, low-rank adaptation (LoRA), and structured pruning.

Bridge the probability and decision-theoretic ideas that connect Bayesian inference, calibrated prediction, and classical statistical learning.

Connect risk minimization, validation protocol, evaluation metrics, and deployment-shift thinking so model selection and measurement line up with generalization.

Fill in the classical classification concepts around margins, kernels, asymmetric costs, and imbalance-aware objectives.

Anchor reinforcement learning, causal inference, and two landmark papers that shaped modern deep learning and transformers.

Move from hypothesis testing and likelihood methods to uncertainty quantification and probabilistic scoring.

Study classical unsupervised, latent-variable, and sequential-state models that still shape modern machine learning.

Connect reliable experimentation to the dynamic-programming and return-estimation ideas at the core of reinforcement learning.

Bridge causal estimands, interventional reasoning, and the landmark sequence-model ideas that led into modern attention and deep learning.