Diagnosing bias vs variance by experiment

Workflow: (1) Plot learning curves: train and validation error vs training set size at multiple N (e.g., 0.1K, 1K, 10K, 100K). (2) Patterns: high bias: both curves plateau at high error with small gap (model underfits data); high variance: train error low, validation error much higher, gap shrinks with more data. (3) Concrete example: say training error = 5%, validation error = 25% at small data. As data doubles 8x, validation drops to 10%. High variance; solution: more data, simpler model, regularization, ensembling. If instead train error rises (5% -> 12%) with more data: high bias; bigger model, better features, longer training. Senior view: this diagnostic is cheap, often >3 effective experiments return a clear story, and it's the first thing to run before throwing compute at a complex model.