Difficulty: Intermediate

Persistent directional movement over time.

AI that ranks patients by urgency.

Restricting updates to safe regions.

AI giving legal advice without authorization.

When a model cannot capture underlying structure, performing poorly on both training and test data.

Neural networks can approximate any continuous function under certain conditions.

Learning structure from unlabeled data, such as discovering groups, compressing representations, or modeling data distributions.

Review process before deployment.

Categorizing AI applications by impact and regulatory risk.

Maximum expected loss under normal conditions.

Expected cumulative reward from a state or state-action pair.

Inferring and aligning with human preferences.

Gradients shrink through layers, slowing learning in early layers; mitigated by ReLU, residuals, normalization.

Error due to sensitivity to fluctuations in the training dataset.

A measure of a model class’s expressive capacity based on its ability to shatter datasets.

A datastore optimized for similarity search over embeddings, enabling semantic retrieval at scale.

Transformer applied to image patches.

The set of tokens a model can represent; impacts efficiency, multilinguality, and handling of rare strings.

Changing speaker characteristics while preserving content.

Detects trigger phrases in audio streams.

Gradually increasing learning rate at training start to avoid divergence.

Methods to set starting weights to preserve signal/gradient scales across layers.