Machine learning

The scalar a model minimizes: cross-entropy for classes and tokens, MSE or Huber for values, each with a stable form and simple gradient.

Pick initial weight variance so activations and gradients keep their scale through depth; He for ReLU, Glorot for tanh, scaled-down residual outputs for transformers.

Standardize activations per batch, token or group with a learned scale so deep nets train fast and stably; RMSNorm in LLMs, BatchNorm in convnets.

Techniques that trade training fit for generalization: weight decay, dropout, early stopping, label smoothing and data mixing, with the exact update rules and defaults.

Train an encoder so two views of the same thing score higher than every other pair in the batch; InfoNCE with a temperature, no labels needed.

Multi-head, multi-query, grouped-query, sliding-window and latent attention; same softmax(QKᵀ/√d)V core, different trade-offs between quality and KV-cache size.

Encode position by rotating pairs of query and key dimensions by position-dependent angles, so attention scores depend only on relative distance; the LLM default.

Replace one FFN with many expert FFNs and a router sending each token to its top-k; parameters scale with E, per-token compute with k.

A learned table mapping discrete ids to dense vectors: gather forward, sparse scatter-add backward, row-wise optimizers to keep huge tables cheap.

Learn a subword vocabulary by repeatedly merging the most frequent adjacent pair; byte-level BPE encodes any input losslessly with 32k-200k tokens.

Turn files on disk into shuffled, sharded, resumable GPU batches fast enough that the accelerator never waits.

Estimate set Jaccard similarity from k min-hashes and find similar pairs with banded LSH buckets; the standard dedup for web crawls, training corpora and sequence collections.

Generate label-preserving variants of each training example on the fly to enlarge the effective dataset and bake in invariances.

Transform each input and its target together with a symmetry group to multiply training data, or average predictions over the group at test time.

Train by contrasting each observed positive with a few sampled negatives instead of normalising over every candidate; the choice of sampler decides bias and quality.

One definition per feature feeds a history store for training and a latest-value store for serving, so training and inference see the same values as of the same moment.

Fit a weighted sum of features to a number (least squares) or a probability (logistic); the baseline every model must beat.

Add small regression trees one at a time, each fitted to the loss gradient; histogram splits over 256 bins make it the strongest tabular default.

Partition points into k clusters by alternating nearest-centroid assignment and centroid averaging; k-means++ seeding makes it dependable.

Find orthogonal directions of maximum variance by an SVD of the centred data; randomized SVD gets the top k components in a few passes.

Count how often each feature occurs per class and score a new item by summed log-likelihoods; trains in one pass, updates and merges by adding counts.

Score each record by how few random axis-aligned splits it takes to isolate it in a small forest of random trees; anomalies are few and different, so they isolate fast.

Group points that are packed densely, label the rest as noise: no k, arbitrary shapes, two parameters (eps, minPts), cost set by the neighbour index.

Linear model for counts, rates, positive amounts and proportions: a linear score pushed through a link, fitted by iteratively reweighted least squares.

Hash each feature name to an index in a fixed-size vector, with a second hash for the sign, so no vocabulary is needed and memory is fixed.

Learn action values by bootstrapping toward r + γ·max Q(s′, ·); DQN adds a replay buffer, a frozen target network and Huber loss.

Raise the log-probability of actions in proportion to their advantage; PPO clips the probability ratio so several epochs on one rollout stay safe.

A policy-value net guides PUCT search; search visit counts and game results become its training targets, so it learns from its own improved play.

A fixed-size ring buffer of past experience, sampled uniformly or by priority, so an online learner trains on decorrelated mini-batches and reuses each item.

Choose among K options repeatedly while learning their payoffs: epsilon-greedy, UCB1 and Thompson sampling keep total regret low where a fixed split or a fixed exploration rate pays forever.

Precision, recall, ROC AUC, average precision, log loss, NDCG and pass@k computed correctly, plus paired confidence intervals for comparing models.

Make predicted probabilities match observed frequencies: measure with ECE, fix post hoc with temperature, Platt or isotonic scaling on held-out data.

Grade open-ended outputs with a language model under a rubric, then correct its known biases by order swapping, voting and calibration against human labels.

Two separate encoders map users/queries and items into one vector space; item vectors are precomputed and searched by nearest neighbour, so recall over millions costs one small forward pass.

Train a scorer on query-grouped data so that the order it induces maximizes NDCG; LambdaMART (boosted trees with NDCG-weighted pair gradients) is the production default.

Learn a short latent vector per row and per column from the observed cells only; a missing cell is predicted by the dot product of its two vectors.

Embedding tables plus a feature-interaction net predict several feedback signals per candidate; a weighted sum of them is the ranking score.

Learn p(action | user, item, context) from hashed categorical features with interaction terms (FM, Wide & Deep, DCN, DLRM), then keep it calibrated, since downstream multiplies it.

A network turns an image into scored, classified boxes in one pass; anchor-based, anchor-free and set-prediction heads differ in how they assign labels and remove duplicates.

Keep the best-scoring candidate in each neighbourhood and drop the rest; greedy NMS, Soft-NMS and 1-D peak picking share one loop.

Turn per-frame detections (camera boxes or radar plots) into persistent IDs by predicting each track, gating, assigning and managing track births and deaths.

Find text lines in an image, read each crop with a CTC or attention model, then correct with a lexicon and score confidence for review routing.

Turn a scan or phone photo into an upright, evenly lit, binarised or normalised grey page so layout analysis and OCR read it reliably.

Rebuild words, lines, columns and reading order from the glyph positions a born-digital PDF stores, and route pages without usable text to OCR.

Decode bytes, tolerate broken markup, drop non-content nodes and score text blocks by length and link density to keep the main run, with headings, metadata and offsets for citation.

Cut documents into token-bounded pieces along structure boundaries, keeping offsets and section paths, so each piece embeds well and can be cited.

Find the character spans of entities in text and give each a type, by labelling every token and decoding a valid label sequence.

Map each mention to a knowledge-base id, or to NIL, by generating candidates from an alias table, scoring them on prior and context, and thresholding.

Assign one or more labels to a text; climb a ladder from regexes to linear models on n-grams, embeddings, fine-tuned encoders and LLM prompts, only as far as the data and budget demand.

Score a text's character n-grams against per-language statistics, after a cheap script check, and return an ISO 639 code with a confidence or "und".

Rewrite messy text into one canonical form for matchers and models, or into spoken words for speech output (and back), keeping a map to the original offsets.

Realise facts as sentences by filling slots in alternative templates, with agreement, referring expressions and seeded variation; every word traces to data or a template.

Turn one utterance into an intent label and typed slot values with a joint model: a sentence-level classifier plus a token-level BIO tagger on a shared encoder.

Keep a structured belief of the user's goal (slot → value, plus requests) and update it every turn from NLU output, the old state and the last system action.

Per-frame speech/non-speech decision with smoothing, from an energy threshold over a tracked noise floor up to small neural nets.

Streaming sequence model that sums all monotone alignments in a frame-by-label lattice and conditions each label on earlier ones.

Alignment-free sequence loss that sums the probability of every frame-to-label alignment with a blank symbol, using a dynamic program.

One encoder-only transformer reads the board as 64 square tokens and outputs move probabilities and a win/draw/loss estimate for a tree search.