Glossary

It reaches the same quality with 30 to 60 percent of the annotations but makes the dataset unusable as a test set.

See also: Annotation, Test set

Adam keeps two extra states per parameter. At seven billion parameters that is 56 gigabytes for the optimiser alone.

See also: Momentum, LoRA

An agent is a loop: think, call a tool, read the result, repeat. Reliability falls exponentially with the number of steps.

See also: Tool call, Prompt injection

Without an inventory no duty assessment is possible, because it is unknown what it would apply to.

See also: Approval process, High risk

The duty falls on providers and deployers alike and is tied to the concrete use, not to generic training.

See also: Deployer, Human oversight

Alignment teaches preferences, not knowledge and not truth. The guideline given to raters is the real specification of the behaviour.

See also: RLHF, DPO

From a few seconds per case for classification to a quarter of an hour for segmentation. The factor between is about 200.

See also: Label, Cohen's kappa

Pseudonymisation is not anonymisation: pseudonymised data remains personal and stays subject to the GDPR.

See also: Pseudonymisation, Differential privacy

Without an approval process, shadow AI appears: tools in use that nobody knows about and nobody owns.

See also: Evidence, AI inventory

An umbrella term, not a single method. In current usage it almost always means systems that learn from data rather than following written rules.

Every token queries all keys and mixes the values accordingly. Cost grows quadratically with length.

See also: Transformer, KV cache

Under Art. 22 GDPR, rights then arise to human intervention, to express a view, and to contest.

See also: Human oversight, GDPR

The backward pass costs about twice the forward pass but needs a multiple of the memory for intermediates.

See also: Chain rule, Gradient

A base model is not an assistant. Instruction fine-tuning and alignment are what turn it into a system that answers rather than continues.

See also: Fine-tuning, Alignment

Ignoring the base rate is the most common mistake when judging AI results. For rare events it decides everything.

See also: Bayes' theorem, Precision

Continuous batching admits a new request as soon as one finishes and often raises throughput five to twentyfold.

See also: Throughput, Latency

How often a test fires when someone is affected has little to do with how likely being affected is when the test fires.

See also: Base rate

Differences below three percentage points on a general benchmark are worthless for a decision.

See also: Contamination, Evaluation set

Dropping a protected attribute does not remove bias, only its measurability.

See also: Fairness, Representativeness

The remedies point in opposite directions. Get the diagnosis wrong and you make things worse.

See also: Variance, Overfitting

The vector derived from a face is itself biometric data, even after the image has been deleted.

See also: Embedding, High risk

BM25 saturates term frequency: the twentieth occurrence adds little over the tenth. That matches reality.

See also: Hybrid search

A calibration is valid for that rig only. Moving a camera invalidates it.

See also: Disparity

The hysteresis step joins broken lines while suppressing isolated noise points.

See also: Sobel, Contour

The product form of the chain rule immediately explains why gradients vanish or explode in deep networks.

See also: Gradient, Backpropagation

Chunking decides RAG quality more than the choice of model. Split at headings, not by character count.

See also: RAG, Embedding

That lets an image be compared with a sentence without any training for that class.

See also: Embedding, Vision transformer

Below 0.6 means the guideline is unclear. A model cannot reach high accuracy on such a task.

See also: Annotation, Label

The colour space is a decision, not a property of the image. For colour filters HSV is nearly always right.

See also: Pixel

It says how precise the measurement is, not how good the model is. An accuracy figure without an interval is a snapshot.

See also: Standard error

A model that solves a known task and fails the same task with different numbers has not understood it but seen it.

See also: Leakage, Benchmark

The window covers the system instruction, history, attached documents and the answer so far. A larger window does not mean everything in it is used equally well.

See also: Token, KV cache

Contours allow counting, measuring and shape checks with no trained model at all.

See also: Canny, Morphology

Depending on the numbers in the matrix, the result traces edges, blurs, or highlights patterns.

See also: Convolutional network, Kernel

A CNN brings two assumptions that hold for images: neighbourhood matters, and a pattern means the same everywhere.

See also: Convolution, Vision transformer

In the EU, training rests on the text and data mining exception; a purely machine-generated work generally enjoys no protection of its own.

See also: Text and data mining

Two texts on the same topic point the same way even if one is ten times longer.

See also: Dot product, Embedding

Cross-entropy includes the entropy of the true distribution as a constant floor; only the remainder depends on the model.

See also: Entropy, KL divergence, Loss function

With several cases per person you must split by group, or the same person appears in training and test.

See also: Leakage, Test set

It is a required document, not a nicety: the AI Act requires exactly this information for high-risk systems.

See also: Representativeness, High risk

It is required for large-scale processing of special categories, systematic monitoring, and profiling with significant effect.

See also: High risk, Fundamental rights impact assessment

In most projects, cleaning data delivers more than changing model, and it benefits every future model too.

See also: Duplicate, Cohen's kappa

Server location alone is not enough: what also matters is whether a parent company in a third country can gain access.

See also: Processing on behalf, Third-country transfer

Deployer duties are lighter than provider duties but they exist: human oversight, information, logging, suitable input data.

See also: Provider, Human oversight

A derivative is purely local information. It says what a tiny step does, not where the path leads.

See also: Gradient, Chain rule

A good descriptor stays the same under rotation and lighting change, enabling matching across images.

See also: Homography, RANSAC

Dice and IoU induce the same ordering and convert into each other, but their values are not comparable.

See also: Segmentation, IoU

Only with such a guarantee and a documented budget can anonymity be argued to a supervisory authority.

See also: Synthetic data, Anonymisation

Unlike a transformer it does not work sequentially but refines the whole result at every step.

See also: Transformer

Depth error grows with the square of distance: a system accurate to a centimetre at two metres is 25 centimetres out at ten.

See also: Calibration

Distillation cuts cost and latency substantially and costs measurable quality. How much depends on the task and must be measured against your own evaluation set.

See also: Inference

The dot product is why similarity search and attention work at all.

See also: Cosine similarity, Vector

Because the optimal policy can be written in closed form through the reward model, that model drops out of the calculation. What remains is stable supervised training.

See also: RLHF, Alignment

The most common failure is not a crash but silent degradation. Without measurement it surfaces weeks later.

See also: Population stability index, Model registry

It usually does not help in large-scale pre-training, because the data is seen only once. In fine-tuning it very much does.

See also: Regularisation

Duplicates are the most common silent source of error: the same case can land in training and test.

See also: Leakage, Label

Computing on the device is a legal advantage before it is a technical one.

See also: Quantisation, Data residency

An embedding is not anonymised data: a sentence embedding partially reconstructs its source text.

See also: Vector, Cosine similarity, Vector database

Translation and speech recognition use it because the whole input is available before output begins. Language models mostly use the decoder alone, because they keep writing on.

See also: Transformer, Attention, Embedding

A die has more entropy than a coin, because more can happen. In models, high output entropy indicates guessing.

See also: Cross-entropy, Perplexity

The AI Act regulates use rather than technology. The same model can be unproblematic or high risk depending on purpose.

See also: High risk, Provider, Deployer

Fifty of your own cases carry more weight than any external leaderboard. The set is versioned and frozen.

See also: Benchmark, Cross-validation

Evidence does not arise retrospectively. What was not logged in operation cannot be evidenced later.

See also: Logging, Model registry

F1 weights both equally. They almost never are: the cost per error type belongs in the metric, not in a footnote.

See also: Precision, Recall

Several common fairness criteria are provably not simultaneously satisfiable. The choice is normative and requires justification.

See also: Bias

Fine-tuning teaches form, not knowledge. For knowledge, retrieval is cheaper, more current and verifiable.

See also: LoRA, RAG

Memory falls from quadratic to linear in sequence length. Without it, contexts above 8,000 tokens would be unaffordable.

See also: Attention, KV cache

float16 overflows at 65,504, bfloat16 does not. Which is why bfloat16 won out for training.

See also: Quantisation

It sits alongside the data protection impact assessment rather than replacing it; the subject matter overlaps only partly.

See also: High risk, Data protection impact assessment

The AI Act does not replace the GDPR. Both apply side by side, with different triggers and different duties.

See also: Legal basis, Purpose limitation

Learning means shifting every parameter a little against the gradient and repeating that millions of times.

See also: Derivative, Learning rate

When buying, memory decides first, bandwidth second and compute only third.

See also: VRAM, Memory bandwidth

Grounding is the most effective lever against invented statements. It consists of retrieval, a citation duty, and a check that the claims are actually supported.

See also: RAG, Hallucination

With 32 query heads and 8 key-value heads the KV cache is a quarter of the size, at barely measurable quality cost.

See also: KV cache, Attention

A language model produces the most probable continuation. Where no information exists, a plausible one still emerges, and plausibility is not truth.

See also: Grounding, Temperature

High risk follows from Annex I for safety components and Annex III for named fields such as employment and education.

See also: EU AI Act, Fundamental rights impact assessment

The histogram reveals exposure problems in a second and is the cheapest quality check there is.

See also: Pixel, Otsu threshold

On a million vectors HNSW finds around 96 percent of the true nearest neighbours in about one millisecond rather than 380.

See also: Vector database, Recall

Four point pairs determine it uniquely. It is the tool that straightens a document photographed at an angle.

See also: Descriptor, RANSAC

Oversight is not a line in a policy. It needs authority, time, competence and a real way to intervene.

See also: Deployer, High risk

Pure vector search fails on case numbers and proper nouns. The combination regularly lifts recall@10 by 5 to 15 points.

See also: RAG, BM25

During inference nothing about the model changes. Per operation it is far cheaper than training, but over a lifetime it often costs more in total.

See also: Latency, Throughput

IoU is stricter than it looks: two visibly overlapping rectangles can land well below the usual 0.5 threshold.

See also: mAP, Non-maximum suppression

In classical image processing a human picks the values; in a CNN they are learned.

See also: Convolution

In RLHF it measures how far a model has drifted from its starting point, and bounds exactly that.

See also: Cross-entropy, RLHF

The cache grows linearly with context length and concurrent requests and overtakes the model itself at long context.

See also: Attention, Context window

Label quality caps achievable model quality. Five percent wrong labels means at most 95 percent accuracy.

See also: Annotation, Cohen's kappa

More than two seconds to the first character reads as a hang, however fast the rest is.

See also: Throughput, Batching

Unlike BatchNorm, LayerNorm is independent of batch size and therefore works at batch one.

See also: Residual connection

Leakage produces excellent measurements and a worthless model. A result that is too good is nearly always a leak.

See also: Cross-validation, Contamination

Too small means training crawls. Too large means the loss explodes within a few steps.

See also: Gradient, Warmup

Without a legal basis, processing is unlawful however useful it is. Consent in employment is vulnerable.

See also: GDPR, Purpose limitation

A complete log of all inputs is tempting and routinely a processing operation with its own legal basis.

See also: Evidence, Retention period

Temperature divides the logits before they become probabilities, which is why it sharpens or flattens the distribution.

See also: Softmax, Temperature

Memory need falls by a factor of five to ten, and the result is a file of a few tens of megabytes.

See also: Fine-tuning, Adam

It is where business priorities enter the model. Choose it unconsciously and you leave that to a library default.

See also: Cross-entropy, Gradient

mAP at 0.5 to 0.95 is stricter than mAP at 0.5 and therefore the usual comparison figure.

See also: IoU, Precision

A 4096 by 4096 matrix has 16.8 million entries and takes about 33 megabytes in 16-bit representation.

See also: Vector, Parameter

For a single request bandwidth is the limit, not compute: the arithmetic units sit over 99 percent idle.

See also: VRAM, Roofline model

Generating a single token is memory-bound. Batching is how that changes.

See also: Memory bandwidth, Batching

That saves compute, not memory: every expert must be in memory. Which makes the design unsuitable for edge devices.

See also: Parameter, Quantisation

A model is not a program in the usual sense. It consists of an architecture and the parameters stored in it, which came out of training.

See also: Parameter, Training

Without a registry you can neither roll back nor evidence which version produced a given output.

See also: Drift, Rollback

Momentum helps leave flat plateaus and saddle points, which are far more common in high dimensions than local minima.

See also: Gradient, Adam

Morphology repairs what thresholding broke: closing holes, removing fringes, separating connections.

See also: Otsu threshold

Images are translated into tokens and placed in the context like text. One image can cost more tokens than a page of text.

See also: Token, Embedding

The NMS threshold decides the counted number and therefore belongs in the operating documentation.

See also: IoU

Dividing out the norm is the difference between how similar and how similar and how long.

See also: Vector

Errors show up as double counting: a broken track becomes two objects.

See also: Optical flow, Non-maximum suppression

At 98 percent character accuracy, a 22-character IBAN is fully correct only 64 percent of the time.

See also: Rectification

Open weights are not the same as open source: training data and training code are usually absent, and the licence may restrict use.

See also: Model

From a single point only motion across the edge can be determined. That is the aperture problem.

See also: Object tracking

Otsu works well when the histogram has two clear humps, and otherwise picks an arbitrary point.

See also: Histogram, Morphology

Recognisable when the training loss falls while the validation loss rises.

See also: Bias and variance, Regularisation

Parameters are adjusted during training and fixed afterwards. Their count sets memory need directly: two bytes per parameter in 16-bit representation.

See also: Model, Quantisation

Perplexity 8 means as uncertain as guessing among eight equally good continuations. It is not comparable across tokenisers.

See also: Cross-entropy, Tokeniser

A colour image is a three-dimensional array of height, width and channels. That is all it is.

See also: Colour space, Histogram

Below 0.1 counts as unremarkable, above 0.2 as a clear shift that should trigger a check.

See also: Drift

For rare events, precision decides usefulness rather than the detection rate.

See also: Recall, Base rate

Prefill fully loads the card and sets time to first token. Decode afterwards barely loads it.

See also: Latency, KV cache

Every language model output is a probability distribution over the whole vocabulary.

See also: Base rate, Bayes' theorem

It requires a contract with the mandatory content of Art. 28(3), including sub-processors, deletion and audit rights.

See also: GDPR, Data residency

It goes beyond data protection: a processing agreement alone is not enough, the provider must be bound and informed.

See also: Processing on behalf, GDPR

This includes social scoring, emotion recognition in the workplace and in education, and biometric categorisation by sensitive attributes.

See also: EU AI Act, Biometric attribute

A prompt is more than a question. It carries role, task, context, examples and constraints on form. Answer quality depends on it more than on the model.

See also: Context window, System prompt

The invariant part must sit at the very front or the cache does not engage. Savings in practice are 30 to 60 percent of input cost.

See also: KV cache, Context window

As soon as a system reads external content, that content can contain an instruction aimed at it. Read content is always data and never an instruction.

See also: Agent, Tool call

Without a library everyone reinvents their own prompts and nobody knows which version is in use where. It belongs under version control like code.

See also: Prompt, System prompt

Substantially modifying a third-party model or offering it under your own name can make you a provider.

See also: Deployer, EU AI Act

Pseudonymising before the model call is the single most effective risk reduction and is usually possible without drawback.

See also: Anonymisation, Purpose limitation

A PUE of 1.3 means a 30 percent surcharge for cooling and losses. It belongs in every consumption figure.

See also: Edge

Data collected to perform a contract is not thereby usable for model training. A change of purpose must be assessed and justified.

See also: Legal basis, Retention period

int8 usually costs under one percent of quality, int4 one to three. On arithmetic tasks the loss is disproportionate.

See also: Floating point, VRAM

Without a quota the busiest user sets the response time for everyone. It belongs in the first version, not the second.

See also: Throughput, Latency

The practical gain is not better language but the ability to name a source and have it verified.

See also: Grounding, Vector database, Chunking

Without RANSAC a single bad match ruins the computed transform entirely.

See also: Homography, Descriptor

In vector search, recall measures how many of the true nearest neighbours an approximate search found.

See also: Precision, F1 score

Rectification typically lifts OCR accuracy on phone photos from below 60 to above 90 percent.

See also: OCR, Homography

The most effective regulariser is more and more varied data. Every algorithmic measure is a substitute for it.

See also: Overfitting, Dropout

Deviations are normal; unnoticed ones are the problem. Art. 10 AI Act requires an assessment for high-risk systems.

See also: Sample, Bias

Once retrieval returns more than ten candidates, a cross-encoder over 50 lifts precision noticeably.

See also: RAG, Recall

Without residuals every layer has to regenerate its output. With them it only has to say what it wants to change.

See also: Transformer, Layer normalisation

Retention duties and erasure duties routinely conflict. Both belong set separately per data category.

See also: Purpose limitation, Processing on behalf

People choose between two answers, a reward model is trained on those preferences, and the policy is optimised against it. DPO achieves similar results without that detour.

See also: Alignment, DPO

With heavily imbalanced classes ROC AUC looks optimistic; the precision-recall curve is more informative.

See also: Precision, Recall

An unrehearsed rollback takes about three times the estimate when it matters.

See also: Model registry

If arithmetic intensity is below the ratio of peak compute to bandwidth, more compute does not help.

See also: Memory bandwidth

Because RoPE encodes relative distances, context can be extended afterwards by stretching the frequencies.

See also: Transformer, Context window

Deliberately collected hard cases improve training and make a test set unusable.

See also: Test set, Representativeness

It turned a research question into an investment decision: the effect of ten times the budget can be estimated without building.

See also: Training, Parameter

Segmentation costs a multiple of detection in annotation and is usually the bottleneck of a project.

See also: Dice coefficient, IoU

Shadow deployment is the only way to measure a new version on real traffic before it has any effect.

See also: Rollback, Model registry

Magnitude and direction of the edge follow from the two directional derivatives, as for a vector.

See also: Canny, Convolution

Softmax is never computed without subtracting the maximum, or the exponent overflows and the result turns NaN.

See also: Temperature, Logit

The method is lossless: the output distribution stays exactly that of the large model.

See also: Batching, Distillation

To double measurement precision you need four times the data. That is the most commercially significant number in statistics.

See also: Variance, Confidence interval

They do not replace a test set. Evaluation is always against real data, and anonymity without a formal guarantee is unproven.

See also: Differential privacy, Data card

The system prompt sets behaviour, tone and limits and is usually invisible to the user. It belongs under version control and documented.

See also: Prompt

At zero the most probable token always wins. A high temperature produces not creativity but spread.

See also: Softmax, Top-p

They engage only with matching formats and dimensions. A width of 4095 rather than 4096 costs noticeably.

See also: Graphics card, Floating point

As soon as the test set is used repeatedly to choose between variants it is a second validation set and its number is too optimistic.

See also: Cross-validation, Evaluation set

Lawful access is a precondition. Outside research the exception applies only where no machine-readable reservation was made.

See also: Copyright

It needs its own basis: an adequacy decision, standard contractual clauses with a transfer impact assessment, or a derogation.

See also: Data residency, Processing on behalf

Latency and throughput trade off: batching raises throughput and slows the individual request.

See also: Latency, Batching

Cost, context length and perplexity are all measured in tokens. German needs roughly 20 to 40 percent more tokens than English for the same content.

See also: Context window, Tokeniser

German needs roughly 20 to 40 percent more tokens for the same content and therefore costs correspondingly more.

See also: Token, Quota

A model is inseparable from its tokeniser. Two models with different vocabularies have no comparable perplexity figures.

See also: Token

The model produces structured text read as a function call. A tool's description is itself a prompt and decides how often the right one is picked.

See also: Agent

Top-p beats top-k because the candidate set adapts to the situation rather than being fixed.

See also: Temperature

Training means making a prediction, measuring the error, shifting every parameter a tiny step in the direction that lowers it, and repeating that millions of times.

See also: Gradient, Loss function

The 2017 breakthrough is not intelligence but parallelisability: the whole sequence is processed at once.

See also: Attention, Residual connection

Under Art. 50 AI Act, providers must label machine-readably and deployers must disclose deepfakes.

See also: EU AI Act, Watermark

Without a spread figure, a comparison of two models cannot be told apart from noise.

See also: Standard error, Bias and variance

Words, sentences and images are translated into vectors so that similarity becomes a question of distance.

See also: Dot product, Embedding

Below about 100,000 vectors an extension to your existing database suffices; a dedicated system pays only above that.

See also: Embedding, HNSW, Recall

Without a CNN's built-in assumptions, a ViT needs more data but scales better.

See also: Convolutional network, Transformer

A seven-billion-parameter model needs 14 gigabytes in float16, 7 in int8 and 3.5 in int4, plus context memory.

See also: Quantisation, KV cache

For transformers, a warmup over 2 to 5 percent of the steps is effectively standard.

See also: Learning rate

A watermark is a duty-of-care measure, not protection: it helps with cooperative providers and not against deliberate misuse.

See also: Transparency duty

The counterpart is few-shot, where two to five examples sit in the prompt. Where a task fails zero-shot, one example is almost always cheaper than fine-tuning.

See also: Prompt, Fine-tuning