AI and machine learning interview questions

Whitening scales components to unit variance, and sometimes decorrelates them fully, so downstream distance models treat axes equally.

Linear PCA can only find linear subspaces.

Feature hashing maps tokens or categories into a fixed-width vector with a hash function, optionally with signed hashes to reduce collisions.

Using a statistic that includes later periods, a target from tomorrow, or a rolling mean that is not shifted, all leak.

Quantile bins let a linear model learn a piecewise-constant nonlinear effect without polynomials.

If the rare class is extremely scarce and poorly represented, a one-class or density model on the majority may beat a two-class learner.

Synthetic or duplicated minority rows created on the full set can appear in both train and val folds.

Focal loss down-weights easy majority examples so the gradient focuses on hard, often minority, cases.

Keep custom transformers importable and pickle or joblib the fitted Pipeline with the same library versions.

sklearn Pipeline expects transformers that implement transform on X only.

memory= on a Pipeline caches fitted transformer outputs so repeated grid points reuse the same scaled matrix.

Random weights that are too large explode activations

Batch norm standardizes layer inputs using batch statistics and learned scale and shift.

The log-softmax plus negative log likelihood is numerically stable because it avoids overflow from raw exp.

A wide enough shallow net with a nonlinear activation can approximate a continuous function on a compact set to any accuracy.

L1 geometry tends to select a single coordinate from a tight group because any of them can explain the same direction.

Label smoothing replaces the hard 1.0 target with a slightly lower value and spreads the leftover mass over other classes.

Augmentation trains the model on transformed copies so it cannot memorize exact pixels or tokens.

You train on a past window, validate on the next stretch, then slide or expand the origin and repeat.

Outer folds estimate the performance of a method that itself contains a search.