Advanced Deep Learning Interview Questions, issue 15, Apr 5, 2026

The Convexity Assumption Trap

Senior ML Engineer interview at Meta, and the interviewer asks:

You’re migrating a legacy continuous prediction model into a multi-class classifier. A junior dev suggests keeping the L2 (MSE) loss for the new Softmax outputs because ‘error is error.’ Why is this guaranteed to break the optimizer in production?

Don’t say: Because Cross-Entropy is meant for probabilities and MSE is for regression.

Many assume the optimizer sees a smooth bowl, but MSE over Softmax outputs produces a jagged, non-convex surface that standard training loops can’t reliably descend.

The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.

Read it on Substack

Get the next one

Free on Substack. Unsubscribe in one click.