Machine Learning and AI

Routing Summary

This folder covers modern machine learning and AI as seen from a statistician’s desk: how transformers and large language models work, how LLMs are turned into reasoning systems and agents, and two places where ML meets classical inference — distribution-free predictive uncertainty (conformal prediction) and probabilistic forecasting. Contains 29 notes in 4 sub-topics, grounded in 16 primary papers (2007–2024).

Sub-topics

Sub-topicNotesCovers
Transformers and LLM Foundations7Scaled dot-product and multi-head attention, the transformer architecture and positional encoding, autoregressive pretraining, Kaplan scaling laws, Chinchilla compute-optimal training, in-context learning — Vaswani 2017, Kaplan 2020, Hoffmann 2022, Brown 2020
LLM Reasoning, Retrieval and Agents8Chain-of-thought prompting, retrieval-augmented generation, ReAct, the tool-use agent loop, RLHF and instruction tuning, reward modeling as a paired-comparison choice model, evaluation and hallucination — Wei 2022, Lewis 2020, Yao 2022, Ouyang 2022
Conformal Prediction7Split conformal and the exchangeability coverage guarantee, conformity scores and adaptive sets, conformalized quantile regression, marginal vs conditional coverage, weighted conformal under covariate shift, conformal intervals for counterfactuals and ITEs — Angelopoulos & Bates 2021, Romano 2019, Tibshirani 2019, Lei & Candès 2020
Probabilistic Forecasting7Proper scoring rules (CRPS, log score, pinball), DeepAR and global neural forecasters, time-series foundation models (Chronos), local vs global models, MinT hierarchical reconciliation, rolling-origin backtesting — Gneiting & Raftery 2007, Salinas 2017, Ansari 2024, Wickramasuriya 2019

Reading Paths

Cross-Cutting Concepts

Sources

All PDFs are in Machine Learning and AI/raw/; each sub-topic index lists its own sources with arXiv ids.

See Also