Domain: Foundations & Theory
Controls the size of parameter updates; too high diverges, too low trains slowly or gets stuck.
Local surrogate explanation method approximating model behavior near a specific input.
Choosing step size along gradient direction.
Optimal control for linear systems with quadratic cost.
Minimum relative to nearby points.
Raw model outputs before converting to probabilities; manipulated during decoding and calibration.
PEFT method injecting trainable low-rank matrices into layers, enabling efficient fine-tuning.
A function measuring prediction error (and sometimes calibration), guiding gradient-based optimization.
An RNN variant using gates to mitigate vanishing gradients and capture longer context.
Stability proven via monotonic decrease of Lyapunov function.
Predicts masked tokens in a sequence, enabling bidirectional context; often used for embeddings rather than generation.
Mechanisms for retaining context across turns/sessions: scratchpads, vector memories, structured stores.
A parameterized mapping from inputs to outputs; includes architecture + learned parameters.
Standardized documentation describing intended use, performance, limitations, data, and ethical considerations.
Optimizes future actions using a model of dynamics.
Central system to store model versions, metadata, approvals, and deployment state.
Reconstructing a model or its capabilities via API queries or leaked artifacts.
Models that process or generate multiple modalities, enabling vision-language tasks, speech, video understanding, etc.
Training objective where the model predicts the next token given previous tokens (causal modeling).
AI subfield dealing with understanding and generating human language, including syntax, semantics, and pragmatics.
Techniques that stabilize and speed training by normalizing activations; LayerNorm is common in Transformers.
Visualization of optimization landscape.
Finding control policies minimizing cumulative cost.
Coordinating tools, models, and steps (retrieval, calls, validation) to deliver reliable end-to-end behavior.
When a model fits noise/idiosyncrasies of training data and performs poorly on unseen data.
Techniques that fine-tune small additional components rather than all weights to reduce compute and storage.
The learned numeric values of a model adjusted during training to minimize a loss function.
Classical controller balancing responsiveness and stability.
Information that can identify an individual (directly or indirectly); requires careful handling and compliance.
Methods for breaking goals into steps; can be classical (A*, STRIPS) or LLM-driven with tool calls.