Five books that fill the gaps 機械学習 tutorials leave

The list came out of a specific failure. During an internal review, the author reported a score improvement on validation data confidently, then was asked whether the gap fell within the range of chance — and realized he had not considered sample size or variance at all. He deflected the question on the spot, and on the train home concluded that what he had built up was thinner than he had assumed. That experience explains the ordering he now insists on: survey methods broadly first to build a map, then solidify the data and math foundation, then descend into algorithm internals and operations. Read that way, he says, the same books went faster. The five titles reflect that arc. Andriy Burkov's 100+ page essence is positioned as a map for people who have already touched several methods, with short explanations that still say what each method is for and where it breaks. The statistics primer by Peter Bruce, Andrew Bruce and Peter Gedeck is where the author says he learned to fall back on resampling — bootstrap and permutation tests — when he lacks confidence in distributional assumptions, and it covers misuse patterns for p-values and significance. The mathematics book by Marc Peter Deisenroth, A. Aldo Faisal and Cheng Soon Ong is rated hard; the author stopped trying to read it cover to cover and switched to looking up notation that tripped him up in papers, saying the eigenvalue and singular value decomposition chapters alone changed his understanding of PCA. Mykel J. Kochenderfer and Tim A. Wheeler's optimization book sits slightly outside machine learning proper, but the author argues that perspective is the most useful when isolating why training will not converge. Practical MLOps by Noah Gift and Alfredo Deza is included for the part that does not end at deployment: monitoring and logging, where the author says the book gave him criteria for deciding what to measure.