Skip to main content
Fanout
Glossary
AI Research

AI glossary

Plain-English definitions for AI and machine learning terms.

  • Activation function

    The nonlinearity applied after each layer's affine map; without it, a network collapses to a single linear function.

  • Attention

    A weighted average of value vectors, with weights computed from query–key similarity, letting each position look at every other.

  • Backpropagation

    The chain rule applied in reverse through a network to compute every parameter's gradient in one pass.

  • Checkpoint

    A saved snapshot of model (and optimizer) state so training can resume or be evaluated later.

  • Context window

    The maximum number of tokens a model can attend over in one forward pass.

  • Embedding

    A learned dense vector standing in for a discrete token, item, or entity.

  • Fine-tuning

    Continuing training of a pretrained model on a narrower dataset to specialize it.

  • Gradient

    The vector of partial derivatives of the loss with respect to the parameters; the direction of steepest increase.

  • Inference

    Using a trained model to produce outputs — as opposed to training it.

  • KV cache

    The stored key/value tensors from earlier tokens so decoding does not recompute them.

  • LoRA

    Low-rank adaptation: fine-tuning by learning small rank-decomposed updates instead of full weight matrices.

  • Loss

    The scalar a training run minimizes, measuring disagreement between predictions and targets.

  • Overfitting

    When a model memorizes training detail that does not generalize to unseen data.

  • Pretraining

    The large, generic first training phase before any task-specific tuning.

  • RLHF

    Reinforcement learning from human feedback: tuning a model against a learned preference model.

  • Softmax

    Turning a vector of scores into a probability distribution over classes.

  • Temperature

    A scaling knob on logits before softmax; higher makes outputs flatter, lower sharper.

  • Token

    The atomic unit of model input — a word piece, character run, or byte span depending on the tokenizer.