AI and machine learning interview questions

Semi-supervised learning mixes a small labeled set with a large unlabeled set.

Self-supervised learning invents a pretext task from the data itself, such as predicting a masked token or an image rotation.

A discriminative model, such as logistic regression, models P(y|x) and is usually stronger for pure prediction.

If training and validation error are both high and close, you mostly have bias or an unsolvable task.

More independent training data usually lowers variance because the fit is less hostage to any one sample.

K equal to 1 is a high-variance, low-bias classifier that can zigzag around every point.

Stronger L2 or smaller trees increase bias by shrinking or simplifying the fit.

Early stopping tracks a validation metric and freezes weights when that metric stops improving.

Extra columns raise capacity and the chance of spurious correlations in a finite sample.

Interpolation means the model can fit the training points, even perfectly.

Leakage is any path that lets information from the evaluation set, or from the future, into training.

Use chronological cuts so training always precedes validation and test.

A stratified split keeps class proportions similar in train, val, and test.

ROC-AUC is the probability that a random positive scores higher than a random negative.

Precision-recall AUC focuses on the positive class and reacts to prevalence.

It counts true positives, false positives, true negatives, and false negatives at a chosen threshold.

Log loss, or binary cross-entropy, penalizes confident wrong probabilities heavily.

R-squared is the fraction of variance in y explained by the model.

Multicollinearity means features move together so the design matrix is nearly singular.

The closed form needs a matrix inverse or solve on p features and can be costly or unstable for huge p.