Difficulty: Intermediate

Flat high-dimensional regions slowing training.

A point where gradient is zero but is neither a max nor min; common in deep nets.

Automated detection/prevention of disallowed outputs (toxicity, self-harm, illegal instruction, etc.).

Software regulated as a medical device.

Stochastic generation strategies that trade determinism for diversity; key knobs include temperature and nucleus sampling.

Empirical laws linking model size, data, compute to performance.

Repeating temporal patterns.

Optimization using curvature information; often expensive at scale.

Methods to protect model/data during inference (e.g., trusted execution environments) from operators/attackers.

Assigning labels per pixel (semantic) or per instance (instance segmentation) to map object boundaries.

Attention where queries/keys/values come from the same sequence, enabling token-to-token interactions.

Models evaluating and improving their own outputs.

Learning from data by constructing “pseudo-labels” (e.g., next-token prediction, masked modeling) without manual annotation.

Retrieval based on embedding similarity rather than keyword overlap, capturing paraphrases and related concepts.

Pixel-wise classification of image regions.

Training with a small labeled dataset plus a larger unlabeled dataset, leveraging assumptions like smoothness/cluster structure.

Ability to correctly detect disease.

Models estimating recidivism risk.

Fine-tuning on (prompt, response) pairs to align a model with instruction-following behaviors.

AI used without governance approval.

Running new model alongside production without user impact.

Feature attribution method grounded in cooperative game theory for explaining predictions in tabular settings.

A narrow minimum often associated with poorer generalization.

Simultaneous Localization and Mapping for robotics.

Converts logits to probabilities by exponentiation and normalization; common in classification and LMs.

Attention mechanisms that reduce quadratic complexity.

Identifying speakers in audio.

Of true negatives, the fraction correctly identified.

Converting audio speech into text, often using encoder-decoder or transducer architectures.

Generating human-like speech from text.