Machine-Learning

Disclosure: this post was written with the assistance of an AI system (Claude), which wrote the analysis code, ran the experiments and drafted the text. The topic, the constraints, the data choices and the final review are the author’s. A gradient check is a finite difference, and a finite difference has an optimal step size: truncation error falls as the step shrinks and cancellation rises like eps/h, so the total is a U and refining past its bottom makes the answer worse — 9....

Disclosure: this post was written with the assistance of an AI system (Claude), which wrote the analysis code, ran the experiments and drafted the text. The topic, the constraints, the data choices and the final review are the author’s. Fit a step on all the rows, then cross-validate, and compare with the same step fitted inside the folds. Standardising leaks +0.05 points on the breast cancer data and mean imputation -0....

Disclosure: this post was written with the assistance of an AI system (Claude), which wrote the analysis code, ran the experiments and drafted the text. The topic, the constraints, the data choices and the final review are the author’s. On 10,240 next-character predictions from the committed model, accuracy is 0.5466 at all seven temperatures from 0.5 to 3, identical to the bit, while ECE runs from 0.026 to 0.375. Between predictions the confidence order moves: at T = 2 Kendall’s tau against T = 1 is 0....

Disclosure: this post was written with the assistance of an AI system (Claude), which wrote the analysis code, ran the experiments and drafted the text. The topic, the constraints, the data choices and the final review are the author’s. The LayerNorm Jacobian is the orthogonal projector off the plane spanned by the all-ones vector and the normalised input, divided by sigma - verified against autodiff to 2e-16, rank exactly d minus 2, and with a flat nonzero spectrum because it really is a projector....

Disclosure: this post was written with the assistance of an AI system (Claude), which wrote the analysis code, ran the experiments and drafted the text. The topic, the constraints, the data choices and the final review are the author’s. Three training runs of the same 816,128-parameter model, evaluated at every checkpoint on 10,240 held-out predictions. Every run starts slightly underconfident; overconfidence begins after one to ten passes over the training text and grows as training loss pulls away from validation loss....

Disclosure: this post was written with the assistance of an AI system (Claude), which wrote the analysis code, ran the experiments and drafted the text. The topic, the constraints, the data choices and the final review are the author’s. On a simulated task whose Bayes error is 0.268, with training sets grown to 25,600 rows, the two-parameter power law a n^-b fitted to the smaller sizes predicted too low an error in all 9 fits, and in 4 of them predicted an error below the Bayes error - 0....

Disclosure: this post was written with the assistance of an AI system (Claude), which wrote the analysis code, ran the experiments and drafted the text. The topic, the constraints, the data choices and the final review are the author’s. The gradient of the expectation is p times loss minus its mean, which is episode 2’s softmax Jacobian applied to the losses and sums to exactly zero. Over 65 tokens it is enumerable, so every estimator has an answer to be wrong against....

Disclosure: this post was written with the assistance of an AI system (Claude), which wrote the analysis code, ran the experiments and drafted the text. The topic, the constraints, the data choices and the final review are the author’s. Over 200 random row splits of 24,576 predictions, realised conformal coverage at alpha = 0.1 has an SD of 0.0040, 1.49 times the Beta spread episode 1 compared it with. That comparison was short a term: the coverage is measured on a finite test half, and with the test half’s binomial variance added the excess is 1....

Disclosure: this post was written with the assistance of an AI system (Claude), which wrote the analysis code, ran the experiments and drafted the text. The topic, the constraints, the data choices and the final review are the author’s. At 2% prevalence, logistic regression trained on an oversampled, SMOTE-augmented or class-weighted copy of the data ranks the test set no better than the plain model - AUC -0.0012 for oversampling, small and consistently negative - while its mean predicted probability goes from 0....

Disclosure: this post was written with the assistance of an AI system (Claude), which wrote the analysis code, ran the experiments and drafted the text. The topic, the constraints, the data choices and the final review are the author’s. The embedding table’s median pairwise distance is 11.214 and sqrt(2) times its median norm is 11.210 - the distance between orthogonal vectors of that length, matched to four figures, with a median pairwise cosine of +0....