About 150 terms, in plain English, tagged by course
Plain English. Course number in brackets.
Activation function [2] — the non-linear squash applied after a neuron's weighted sum. Without one, a whole network collapses into a single linear equation.
Adapter [10] — a small trainable module inserted into a frozen pretrained model.
AdamW [9] — the standard transformer optimiser; Adam with weight decay applied separately from the gradient.
Advantage (Â) [11] — how much better an action was than expected. Drives the PPO update.
Agent [12] — an LLM that decides what actions to take, calls tools, reads the results, and iterates. Opposed to a chain, where you decide the steps.
Alignment [11] — making a model behave the way humans want: instruction tuning plus preference tuning.
Attention [9] — every token looks at every other token and takes a weighted blend of what it finds. softmax(QKᵀ/√d_k)V.
Attention mask [7] — tells the model which positions are real and which are padding. Omitting it corrupts results silently.
Autoencoder [2,3] — encoder compresses to a bottleneck, decoder reconstructs. Unsupervised; the label is the input.
Autograd [4] — PyTorch's automatic differentiation. Records operations, then computes gradients on .backward().
AUC [1] — area under the ROC curve. 1.0 perfect, 0.5 random. Threshold-independent.
Backpropagation [2] — the chain rule applied backwards through layers to get every weight's gradient.
Bagging [1] — train many models on bootstrap resamples and average. Reduces variance. Parallel.
Batch normalisation [5] — normalise a layer's pre-activations across the batch, then apply learned scale γ and shift β.
Beam search [8] — keep the k best partial sequences while decoding.
BERT [7,9] — encoder-only, bidirectional, pretrained with masked-LM. For understanding, not generation.
Bias (parameter) [2] — the constant added to a weighted sum; lets a neuron shift its threshold.
Bias (error) [1] — error from a model too simple for the problem. Underfitting.
Boosting [1] — train models sequentially, each fixing the last one's errors. Reduces bias. Sequential.
BPE [7] — Byte-Pair Encoding. Subword tokenisation by repeatedly merging the most frequent adjacent pair.
Bradley-Terry [11] — the model behind preference learning: P(A ≻ B) = σ(score_A − score_B).
Catastrophic forgetting [10] — fine-tuning at too high a learning rate erases pretrained knowledge.
Causal mask [9] — blocks a position from attending to later positions. Required for autoregressive generation.
Chain (LangChain) [12] — a fixed sequence of steps. Predictable, cheap, debuggable.
Chain-of-Thought [12] — prompt the model to reason step by step. Works because it buys more computation before the answer.
Chunking [12,13] — splitting documents for RAG. 1000 chars / 200 overlap is the default; the parameter that most determines RAG quality.
[CLS] [7,9] — the special first token whose final representation stands for the whole sequence.
Collate function [7] — assembles a batch; where padding belongs.
Contextual embedding [8,9] — a word's vector depends on its sentence. What BERT gives you and Word2Vec doesn't.
Convolution [2,5] — slide a learned kernel over a grid to produce a feature map. Parameter sharing is the whole advantage.
Cross-attention [9] — decoder queries against encoder keys and values. Where translation happens.
Cross-entropy [1,2] — the classification loss. Equals negative log-likelihood.
Cross-validation [1] — split k ways, train k times, average. A stable generalisation estimate.
Data leakage [1] — information from outside the training set entering training. Looks like excellent results; collapses in production. Pipelines prevent it.
DBSCAN [1] — density-based clustering. No k needed; finds arbitrary shapes; labels outliers as noise.
Decoder-only [7,9] — GPT-style. Causal attention, next-token prediction, generates.
Diffusion model [3,7] — learn to reverse a noising process. Current best for images. Stable to train, slow to sample.
Dropout [5] — randomly zero neurons during training. Prevents co-adaptation. Disabled by model.eval().
DPO [11] — Direct Preference Optimization. Trains on preference pairs directly, no reward model and no RL. Two models instead of PPO's four.
Early stopping [2] — halt when validation loss stops improving. The cheapest regularisation.
Embedding [8] — a dense learned vector representing a token or document. Similar meanings land nearby.
EmbeddingBag [8] — averages a variable-length sequence's embeddings in one op, no padding needed.
Encoder-only [7,9] — BERT-style. Bidirectional, understands, doesn't generate.
Encoder-decoder [8,9] — T5/BART-style. Reads one sequence, writes another. For translation and summarisation.
Epoch [2] — one full pass over the training data.
Experience replay [3] — store past transitions and train on random samples. Breaks temporal correlation.
Exposure bias [8] — a teacher-forced model has never seen its own mistakes, so errors compound at inference.
FastText [8] — embeds character n-grams, so it handles unseen words and typos.
Few-shot [12] — put examples in the prompt. No weight updates.
Fine-tuning [10] — continue training a pretrained model on your task. Teaches behaviour and style, not facts.
Flash Attention [9] — an I/O-efficient exact attention kernel. Same maths, far less memory traffic.
Forward propagation [2] — input to output through the layers.
GAN [3,7] — generator versus discriminator. Sharp output, unstable training, mode collapse.
Gradient accumulation [9] — accumulate gradients over several small batches to simulate a large one.
Gradient descent [2] — step downhill on the loss surface. w := w − α·∇J.
Gradio [13] — Python library for quick web UIs. What the capstone bot runs on.
Greedy decoding [8] — always take the most likely next token. Fast, repetitive.
Hallucination [7] — fluent, confident, false output. Inherent to next-token prediction. RAG is the architectural answer.
HDBSCAN [1] — hierarchical DBSCAN. Handles varying density, removes the eps hyperparameter.
He initialisation [5] — variance 2/n_in. Pairs with ReLU.
Hugging Face [10] — the pretrained-model ecosystem. Sits on top of PyTorch.
Hybrid search [12] — dense semantic search combined with sparse BM25 keyword search.
In-context learning [12] — learning from prompt examples with no weight updates. Emergent at scale.
Instruction tuning / SFT [11] — supervised fine-tuning on (instruction, response) pairs. Teaches the model to answer rather than continue.
KL divergence [11] — how far one distribution is from another. In RLHF it's the leash keeping the policy near the reference.
KNN [1] — classify by majority vote of the k nearest training points. No training; needs scaling.
LangChain [12,13] — framework for composing LLM applications. prompt | llm | parser.
LayerNorm [3,9] — normalise per sample across features. Batch-size independent. Standard in transformers.
LCEL [12] — LangChain Expression Language; the | pipe syntax.
LoRA [10] — Low-Rank Adaptation. Learn W₀ + BA with tiny B and A. B starts at zero. <1% of parameters, mergeable for zero inference cost.
Lost in the middle [13] — models attend better to the start and end of a long context than the middle.
Masked LM [9] — BERT's objective. Mask 15% of tokens (80/10/10 split) and predict them.
Mini-batch gradient descent [4] — update on batches of 32–128. The practical default.
MMR [13] — Maximal Marginal Relevance. Balances retrieval relevance against diversity.
Mode collapse [3] — a GAN generator produces only one output that reliably fools the discriminator.
Momentum [5] — accumulate a velocity vector. Rolls through ravines, saddles and plateaus. ρ = 0.9.
Multi-head attention [9] — several attention operations in parallel, each learning a different relationship type.
Negative sampling [8] — replace a full-vocabulary softmax with k+1 binary decisions. Made Word2Vec trainable.
NF4 [10] — 4-bit NormalFloat. QLoRA's quantisation format, optimal for normally-distributed weights.
NSP [9] — Next Sentence Prediction, BERT's second objective. Later shown to add little.
One-hot [8] — one 1, rest 0. Huge, sparse, semantically useless — all pairs equidistant.
Overfitting [1] — memorising the training set. Train loss falls, validation loss rises.
Padding mask [7,9] — hides [PAD] positions from attention. Distinct from the causal mask.
PCA [1] — linear projection onto the directions of maximum variance. Scale first.
PEFT [10] — Parameter-Efficient Fine-Tuning. Adapters, LoRA, soft prompts.
Perplexity [8] — exp(cross-entropy). Effective number of choices per token. Lower is better.
Pipeline (sklearn) [1] — chains preprocessing and model so preprocessing fits only on training folds. The anti-leakage tool.
Positional encoding [9] — position information added to embeddings, because attention is order-blind.
PPO [11] — Proximal Policy Optimization. Clips the probability ratio to ±ε so no update wrecks the policy. Needs four models.
Precision [1] — of what I flagged, how much was right. TP/(TP+FP).
Prompt engineering [12] — designing prompts: role, task, context, examples, output format.
Quantisation [10] — store weights in fewer bits. FP32 → FP16 → INT8 → NF4.
QLoRA [10] — 4-bit frozen base + 16-bit LoRA adapters. NF4 + double quantisation + paged optimizers.
RAG [12,13] — Retrieval-Augmented Generation. Retrieve documents, put them in the prompt. Supplies knowledge.
ReAct [12] — Reason + Act. The agent loop: Thought → Action → Observation → repeat.
Recall [1] — of what was really there, how much I caught. TP/(TP+FN).
Re-ranking [12] — retrieve many cheaply, then rescore with a cross-encoder and keep the best.
Regularisation [1] — penalise complexity to reduce overfitting. L1 zeroes features, L2 shrinks.
ReLU [2] — max(0,z). Derivative 1 for positive input. The fix for vanishing gradients.
Residual connection [5,9] — output = F(x) + x. Gives gradients a direct path; makes deep nets trainable.
Reward hacking [11] — the policy exploits the reward model rather than satisfying the real preference. Goodhart's law.
Reward model [11] — LM with a scalar head, trained on human comparisons. Only score differences matter.
RLHF [11] — Reinforcement Learning from Human Feedback. reward − β·KL.
RoPE [9] — rotary positional embeddings. What most current LLMs use.
Self-attention [9] — Q, K and V all from the same sequence.
SentencePiece [7] — language-agnostic subword tokenisation. Marks word starts with ▁.
Sigmoid [2] — 1/(1+e⁻ᶻ). Binary output layer, or multi-label. Derivative caps at 0.25.
Silhouette score [1] — cluster quality without labels. Range −1 to 1.
Skip-gram [8] — Word2Vec variant: predict context from the word. Better on rare words. (CBOW is the reverse.)
Softmax [2,5] — turns logits into a probability distribution over mutually exclusive classes.
Soft prompts [10] — learned continuous "virtual tokens" prepended to the input. Cheapest PEFT.
state_dict [4] — PyTorch's parameter dictionary. Save this, not the model object.
Stuff / map_reduce / refine / map_rerank [13] — LangChain chain_type strategies for combining retrieved chunks.
SVM [1] — maximum-margin classifier; kernel trick for non-linear boundaries. Needs scaling.
t-SNE [1] — non-linear 2D visualisation. Preserves local neighbourhoods only. Never use as model input.
Teacher forcing [8] — feed the decoder true previous tokens during training. Fast; causes exposure bias.
Temperature [8,12] — divides logits before softmax. 0 = deterministic, >1 = creative.
Tensor [4] — n-dimensional array with GPU support and gradient tracking.
TF-IDF [8] — term frequency weighted down by document frequency. Rewards distinctive terms.
Tokenisation [7] — splitting text into model units. Subword (BPE/WordPiece/SentencePiece) is standard.
Top-p / nucleus sampling [8] — sample from the smallest token set whose cumulative probability exceeds p. The modern default.
Transfer learning [3] — reuse a pretrained model's features. Freeze base → train head → fine-tune top at low LR.
Transformer [9] — attention-only architecture. Parallel, long-range, O(n²) in sequence length.
Transpose convolution [3] — learned upsampling. Watch for checkerboard artifacts.
UMAP [1] — non-linear dimension reduction. Faster than t-SNE, keeps more global structure, can transform new data.
VAE [3,7] — encodes a distribution rather than a point, so it can generate. Loss = reconstruction + KL.
Vanishing gradient [2] — activation derivatives below 1 multiplied through depth drive early-layer gradients to zero.
Variance (error) [1] — sensitivity to the training sample. Overfitting.
Vector database [12,13] — stores embeddings with approximate nearest-neighbour search. Chroma, FAISS, Pinecone, pgvector.
Vision Transformer (ViT) [6] — image as a sequence of patches. Weak inductive bias; needs scale or pretraining.
watsonx [7,13] — IBM's AI platform. .governance for monitoring and compliance.
Word2Vec [8] — static word embeddings from context prediction. CBOW and skip-gram.
WordPiece [7] — BERT's subword tokeniser. Marks continuations with ##.
Xavier/Glorot initialisation [5] — variance 2/(n_in+n_out). Pairs with tanh/sigmoid.
Zero-shot [12] — instruction only, no examples.