AI and machine learning interview questions

Batch learning fits on a fixed dataset and is redeployed when you retrain.

Inductive learning builds a model that should work on any future point from the same distribution.

For squared loss, expected error equals bias squared plus variance plus irreducible noise.

Double descent is the pattern where test error falls, rises near the interpolation point, then falls again as models grow even larger.

Averaging uncorrelated errors shrinks variance roughly like 1 over the number of models if bias is unchanged.

Overfitting is a train-versus-val gap on data drawn from the same process.

A single val set used for many hyperparameter trials becomes an optimistic selector.

Keep a final test set untouched.

Fitting StandardScaler, PCA, or a target encoder on train plus val before splitting leaks holdout information into features.

Multiple rows from one person share physiology, devices, and labeling quirks.

Discrimination is ranking: do positives score higher than negatives?

Macro F1 averages F1 per class equally, so rare classes count.

Matthews correlation uses all four confusion-matrix cells and stays near zero for majority-class dummy models.

Assign a rupee or dollar cost to false positives and false negatives, then minimize expected cost on validation.

Ordinary least squares needs X transpose X to be invertible, which fails when p exceeds n or columns are collinear.

Squared loss explodes with outliers and estimates a mean.

Forming X transpose X squares the condition number and can lose precision.

Complete separation means a linear rule already classifies the training set perfectly, so maximum likelihood wants infinite coefficients.

Add a few domain interactions when you need a still-interpretable linear log-odds model.