Glossary · Category

Information Theory, Statistics & Core Math

18 plain-English definitions.

Backpropagation: The algorithm that computes gradients of the loss with respect to every model parameter by applying the chain rule backward through the network.
Bias-variance tradeoff: The fundamental machine-learning tradeoff between a model being too simple to capture patterns (high bias) versus too sensitive to training-data noise (high variance).
Bootstrap resampling: A statistical technique that repeatedly resamples an evaluation set with replacement to estimate the variance/confidence interval of a reported metric.
Confidence interval: A statistical range expressing the uncertainty around a measured metric (e.g., benchmark accuracy) given a finite evaluation sample size.
Cross-entropy loss: The standard training loss for language models, measuring the divergence between the model's predicted next-token distribution and the true (one-hot) target token.
Double descent: An empirical phenomenon where test error decreases, then increases, then decreases again as model size or training time grows past the classical overfitting point.
Dropout: A regularization technique that randomly zeroes a fraction of activations during training to prevent co-adaptation and overfitting.
Entropy (information theory): A measure of the uncertainty or "surprise" in a probability distribution; used to analyze token-prediction confidence and some adaptive sampling methods.
Gradient descent: The core optimization algorithm that iteratively adjusts model parameters in the direction that reduces training loss.
Grokking: A phenomenon where a model suddenly transitions from memorization to true generalization long after training loss has plateaued, observed in some small-scale algorithmic tasks.
KL divergence (Kullback-Leibler divergence): A measure of how one probability distribution differs from a reference distribution, used in RLHF/DPO objectives to constrain how far a fine-tuned policy drifts from its reference model.
Log-likelihood: The logarithm of the probability a model assigns to a given sequence of tokens; the basis for perplexity and cross-entropy loss computation.
Lottery ticket hypothesis: The finding that a randomly initialized dense network contains a much smaller subnetwork that, trained in isolation, can match the full network's performance.
Low-rank approximation: Representing a large matrix as the product of two smaller matrices with a much lower inner dimension, the mathematical basis of LoRA and other compression techniques.
Singular value decomposition (SVD): A matrix factorization technique used in some model compression and low-rank adaptation methods to identify the most important directions of variation in a weight matrix.
Softmax function: A function that converts a vector of raw scores (logits) into a probability distribution by exponentiating and normalizing, used at the output layer of a language model and inside attention.
Statistical significance testing: Formal statistical methods used to determine whether an observed difference between two models' benchmark scores is likely real or due to chance.
Weight decay: A regularization technique that shrinks model weights toward zero during training to reduce overfitting.