Alterna AI dictionary

Content and key concepts

AI / Artificial Intelligence
An area of methods and systems for recognizing, planning, finding solutions and other intellectual tasks, not synonymous with neurosnets.
ML / Machine training
Methods of behaviour formation based on data and experience regarding the purpose and quality criterion.
DL / Deep Learning
The ML approach with the compositions of the learning changes and multilevel representations.
Model / Model
In this chapter, parameterized exit transformation. Architecture and parameter values differ.
Options / Weights
The numbers that set the selected function of the model; may be trained or assigned manually. Not current activations.
Dataset
A set of examples for learning or evaluation. Not the basis that each forward is required to reread.
Activation/Activation
The intermediate numerical result of a specific calculation at the current entrance.
Training
In the procedure under consideration: Calculation of the loss and gradient and subsequent updating of the parameters.
Inference
Here: application of the fixed parameters model. Do not confuse the probability conclusion in other contexts.
Token / Token
Disaggregated text coding unit. Could be a word, part of word, symbol, or byte depending on tokenizer.
Embedding
The vector representation; embedding lokup selects the line of the ID table. Coordinates are not required to have human names.
Attention
Query/key coefficients and values blending. The coefficients depend on input and are not identical with projection parameters.
Logits
Unnormal physical evaluations of candidates, not likely.
Softmax
Transforming z estimates into distribution: exp(z_j) divided by the sum of the exhibit. Subtracting the total maximum improves sustainability.
Context / Context
Data transmitted for the current model call: request, instructions, selected history and external information.
RAG
Generation using recovered external information. Searching on request does not per se update the weight of the generator.
Application Memory
External records stored between calls. The model is influenced when the application uses them to assemble the input.
Tool
External perpetrator of a limited operation: The right and arguments are checked by ringime and not by the text of the document.
Skill
Within the course: a set of instructions/work descriptions. The term does not automatically change weights or rights.
Agent runtime
The programme, which manages the call cycle of the model, tools and observations with restrictions and stoppage conditions.
Vector
The vector space element. Once the base is selected, it is recorded by the orderly coordinates.
Base
Linean independent space-producing set. Coordinates depend on the base.
Norma
Non-negative length with uniformity and inequality of triangle; zero only on the zero vector.
A scalar work
Amount of paired coordinates; scalar result.
Cosine similarity
Dot divided into the Euclidic rules of non-zero arguments.
Orthogonal projection
The nearest vector of the chosen subspace in the Euclidean metric is his orthogonal.
Rang
Size of linear transformation image; numerical rank depends on tolerance.
SVD
Degradation of A=U\VT to orthogonal bases and non-negative singular scales.
Gradient
Private scalar derivative vector; local growth relative to selected coordinates/meters.
Jacobian
The J_ij matrix is based on the shape [exit, exit].
Hessian
A matrix of the second derivatives of the scalar function describing the local curve.
Likelihood
Probability/density of fixed data as a function of parameters; not posterior without prior.
Enthropia
Average -log p by distributionp; nats for ln and bits for log2.
Cross-entropy
Average -logq by target. CE(p,q)=H(p)+KL(p
KL divergence
Average log(p/q) pp; not symmetrical, not metric.
Perplexity
An exhibitor of medium token log-loss in the bow, or 2 degrees log-loss in.
Adam
Adaptive optimizer with EMA first/second moment and correction of initial displacement.
Generalization
The quality of new data is different from minimizing the error of the already used training set.
Precision
TP/(TP+FP): the proportion of true assumptions among the predicted assumptions; at zero denominator, the convenation is required.
Recall
TP/(TP+FN): the percentage of positions found among real positions.
Centroid K-means
The arithmetic mean of the designated non-empty cluster points at squared Euclidean objective.
PCA
Orthogonal directions for maximum dispersion of centralized data; not supervised importance for Target.
Activation function
Preactivation conversion is often non-linear and elemental. Do not confuse with a specific activation value.
MLP
The song is composed of a dense affine layers and hidden activations with a separate output layer.
LayerNorm
Standardizing by the coordinates of the selected representation of one example, usually followed by a training affine conversion.
Residual path
The addition of x+F(x) with matching Shapes; creates a straight path, but does not guarantee a non-zero total.
Automatic differentiation
Computing derivatives through local rules and chain rule; not financial defaults or optimizer update.
Topological order
The order of the DAG knots, in which each parent precedes its dependent knot.
Floating point
Presentation via sign, extonent and significand with final accuracy and special coding.
Subnormal
The number with minimum encoded extonent and without a hidden unit; provides a gradual underflow in its supporting mode.
FMA
Multiplies with one final rounding; usually two FLOPs.
Arithmetic intensity
The ratio of arithmetical operations to traffic bytes in the chosen memory model.
Roofline
It is not the measurement of the program that counts thrughput as a mini(compute peak,bandwidth×intency).
KV-cache
Cache keys/valuses of previously processed currents with auto-regressive inference; not model parameters.
Aliasing
The indistinguishability of different reference frequencies/lamps after the chosen discreteization.
Swamp
Linear operation y[n]=h[j]x[n-j]; in sliding-window the shape includes a core revolution.
Cross-correlation
The material rolling sum of the works without a core perversion in the Windows convention adopted here.
DFT
FFT is a family of fast algorithms for his computation.
Receptive field
An area of reference entry capable of influencing the selected output element via a computational graph.
Byte-level BPE
The trained association of frequent neighbors byte-token paires; specific normalitation, pre-tokenization and Franks are part of the tokenizer contract.
Grapheme cluster
Unicode code points group, perceived as a single sign; not necessarily token model.
Skip-gram
The task of predicting the content tokens from the central token with the introductory/outset presentations that are taught.
Contextual state
Activation of the item after processing the environmental sequence; not the new line of the line weights embedding.
Causal mask
Restriction to current and previous entries applied prior to softmax.
Multi-head attention
Independent attributions in projectati