Glossar

Representing a scene with many 3D Gaussian ellipsoids for fast rendering

A high-throughput parallel unit such as a GPU or TPU

What the agent can do; either discrete or continuous

Expected discounted return after forcing the first action to be a

A function that applies a nonlinear transform to the weighted sum

The policy network and the value network: one acts, one scores

How much better an action is than the average at that state

Making model behaviour match human intent and values

Finding the few samples that deviate from the bulk distribution

Letting a framework compute exact gradients automatically, not by numerical approximation

How many samples estimate the gradient per step

How far the model’s average prediction departs from the true regularity

A learnable offset applied to the threshold

How many bits encode each channel; 8 bits give 256 levels

The low-dimensional layer holding the latent code, limiting its bandwidth

A rectangle represented as (x, y, w, h) or corner points

Byte-Pair Encoding: bottom-up merging of frequent symbol pairs

Backpropagation through time after unrolling

Rapid loss of old abilities while learning a new task

The derivative of a composition is the product of the local derivatives

Making the model write out intermediate reasoning steps

A distinct measurement at the same location, such as R/G/B or alpha

Splitting long documents into retrievable pieces

Extrapolating between conditional and unconditional predictions to control prompt fidelity

Grouping samples by similarity (k-means, hierarchical clustering)

A coordinate system for colour values, such as sRGB, HSV or Lab

A computation expressed as nodes and directed edges over which derivatives propagate

A cross-tabulation of true versus predicted classes

Re-forming the batch every step to keep the GPU busy

Learning representations by pulling positives together and pushing negatives apart

A loss that pulls same-class embeddings together and pushes different-class ones apart

A bypass network guiding structure from a condition map

The alignment of two vector directions, from −1 to 1

Query from one sequence, Key/Value from another

The attention mechanism letting image features query text vectors

The information needed to encode data from P using distribution Q

Negative log-probability of the correct class; the default classification loss

The standard objective for classification training

Deterministic sampling achieving comparable quality in a few dozen steps

Discrete Markov diffusion, typically needing a thousand sampling steps

Deeper networks with higher training error, and not from overfitting

Estimating the probability distribution the data follows

The number of entries in a vector

Compressing high-dimensional data to fewer dimensions while preserving structure (PCA, t-SNE, UMAP)

Between 0 and 1; how much future rewards are valued

The network judging real versus fake, serving as the loss

The modern counterexample where test error falls again past the interpolation point

Direct preference optimisation without an explicit reward model

Randomly silencing units during training to prevent co-adaptation

A neuron stuck in the negative region with zero gradient, no longer updating

Halting training before validation loss turns upward

A vector whose direction is unchanged by the map, and the factor by which it is scaled

A lower bound on the log-likelihood: the reconstruction term minus the KL term; a VAE’s actual objective

The layer, or its output, that maps a discrete object into a continuous vector

A model that encodes text into vectors

The model’s average loss on the training samples

When the input shifts, the output shifts accordingly rather than changing

The total probability of the data across all hypotheses; it normalises the result

Store past transitions and sample randomly to break correlation

A model that writes down or approximates p(x), e.g. autoregressive or diffusion

The harmonic mean of precision and recall

Fréchet distance between generated and real distributions in Inception feature space; lower is better

The model emitting structured arguments to invoke an external function

Using 0–1 coefficients from Sigmoid to control how much information passes

Performance on data the model has not seen

The network mapping noise to samples

The magnitude and stability of gradients as they propagate layer by layer

Checks and constraints bounding what an agent may do

A continuously updated "summary so far" vector

Requires the correct class to win by a margin; the heart of the SVM

The set of all functions the model can represent

The path in a residual connection that adds the input straight back to the output

A model that offers only a sampler, not a probability, e.g. a GAN

Solving a task from prompt examples without updating parameters

The standard contrastive loss; essentially a multi-class cross-entropy

Element-wise product summed over entries; the numerator of cosine similarity

The feature vector fed to the model

The shifting distribution of inputs to later layers during training

The ratio of the intersection to the union of two boxes

The unavoidable error floor caused by label noise

Inducing a model past its safety training

A set of learnable weights that slides over the input

The small weight matrix that is learned

Cross-entropy minus true entropy; non-negative and asymmetric

Measures how far the encoded distribution deviates from a standard normal; acts as a regulariser

Penalises divergence from the reference policy to prevent degeneration

Training a small model on a large model’s soft outputs

Caching past tokens’ keys and values to avoid recomputation

The correct output for each sample; the source of supervision

The low-dimensional representation space produced by the autoencoder

How far each step moves

The probability of the observed data under given parameters

The probability of observed data given that the hypothesis is true

A hyperplane exists that separates the two classes perfectly

Turns the gradient of an expectation into a weighted sum of log-probabilities

Influence between elements far apart in a sequence

Low-rank adapters training only a tiny number of new parameters

Low-rank adaptation increments for low-cost customisation

The high-dimensional terrain of loss values over parameter space

Approximating a large matrix by a product of two smaller ones

Mean average precision across classes and IoU thresholds

Hide random words and recover them, a bidirectional objective

An algorithm that evaluates moves via sampled rollouts to guide search

The average squared difference between prediction and label; the default regression loss

The mean of per-class IoU, the primary segmentation metric

When a generator covers only a few modes of the data distribution

The generator covers few modes and loses diversity

The simplest sequential model: unknown reward distributions, one pull per round

Several attentions in parallel, each learning a different focus

Replacing full-vocabulary softmax with a few random negatives

A neural network representing a scene’s radiance field for novel-view synthesis

Non-maximum suppression, removing duplicate boxes

The timetable of noise added per step, described by βₜ or ᾱₜ

A function measuring a vector’s "length"; L2 is the common choice

The behaviour policy may differ from the policy being learned

A sparse vector with a single 1; distinct words are fully orthogonal

A word absent from the vocabulary, spelled out from subwords

Adding zeros at the border to control output size

The exponential of the cross-entropy; the effective number of options the model hesitates among per step

Placing different layers on different devices and filling bubbles with micro-batches

The smallest sampling unit of an image, carrying one or more channel values

A mapping from states to actions, or to a distribution over actions

An explicit order signal, sinusoidal or RoPE

The updated degree of belief after incorporating the evidence

A layout placing normalisation before the convolution, which trains more stably

Placing layer norm before each sublayer for stability

Precision asks how many alerts are real; recall asks how many real cases were caught

Two candidate outputs for one input plus the human’s choice between them

The degree of belief in a hypothesis before seeing data

The "thickness" of probability for a continuous variable; its integral over an interval is the probability

The MLP the contrastive loss is applied to, usually discarded after training

Smuggling malicious instructions as data for the model to follow

Removing low-impact weights or whole structures

Representing float weights and activations with low-bit integers

The three vector roles: what you seek, what is on offer, what is carried

A function mapping outcomes of a random experiment to numbers

The number of independent directions the map actually spans; at most rows or columns

A prompting paradigm alternating reasoning and action

The region of the original input that a given output covers

The input region that one output pixel depends on

Actively hunting for failure and misuse paths

The gap between realised cumulative reward and always picking the best arm

Writing sampling as a deterministic transform plus external noise so gradients flow

Rescoring candidate passages with a more accurate model

Adding the input past a sublayer to ease vanishing gradients in depth

Exploiting the proxy reward instead of genuinely completing the task

A model fitting human preferences and emitting a differentiable score

Area under the ROC curve, measuring ranking ability across all thresholds

Rotary Position Embedding: relative position with better extrapolation

A function whose derivative tends to 0 at the extremes, blocking gradients

Attention whose Q, K and V all come from one sequence

The information of a single event, −log p

Generating training data by having an agent play against its past selves

Labels manufactured from the data itself, no manual annotation

Class → class + instance → the two unified

A subword toolkit that runs directly on the character/byte stream

Supervised fine-tuning on instruction–response pairs

Approximating the full gradient with a mini-batch

The average information, or uncertainty, of a random variable

Writing any matrix as the product "rotate · stretch · rotate"

Routing shallow high-resolution features into deep layers to preserve boundaries

A training objective that predicts surrounding words from the centre

Turns a set of real scores into a probability distribution summing to 1

A local unit spanning frames in video, used to model motion

A small model drafts and the large model verifies in parallel to speed up generation

The variables describing the present situation; must satisfy the Markov property

Expected discounted return from s under policy π

How many pixels the window jumps each step

A slowly updated copy of the network providing stable bootstrap targets

The gap between the fresh target and the old estimate

Splitting a single layer’s large matrices across devices

Sample from the posterior and pick the max, auto-directing exploration to uncertainty

Time from sending a request to receiving the first token

One external capability an agent may invoke

Whether the top prediction / top five include the true label

Pre-train on a large dataset, then fine-tune on a small task

Flip a matrix across its diagonal so rows become columns

An upsampling operation common in segmentation decoders

Bounds how far the new policy may drift from the old

Gradients shrinking exponentially as they are multiplied across layers

How sensitive the model is to perturbations of the training set

A store providing nearest-neighbour search over high-dimensional vectors

An architecture that applies a Transformer to image patches

The fixed set of all tokens and their indices

The component that turns acoustic features back into a waveform

An earth-mover distance between distributions, better behaved for training than JS divergence

How strongly an input influences the output; may be positive or negative

Adding a squared-weight penalty to the loss to suppress large weights

Reusing one set of weights across all spatial positions

Sharding optimiser states, gradients and parameters to cut per-device memory

Outputs symmetric about 0, which aids optimisation

Classifying directly with text prompts, without fine-tuning

Explore at random with probability ε, exploit the current best otherwise