| AI ⊃ ML ⊃ deep learning | Each names a smaller field inside the one before | AI, ML, deep learning |
| Rules vs learning | Programming writes the rules; ML learns them from data | AI, ML, deep learning |
| Supervised | Labelled examples: classification (a class) or regression (a number) | Supervised learning |
| Unsupervised | No labels: clustering, dimensionality reduction, anomalies | Unsupervised learning |
| PCA | New columns ordered by variance; keep the first few. Standardise first | Dimensionality reduction |
| Reinforcement | An agent learns from rewards its actions earn | Reinforcement learning |
| Self-supervised | The data supplies its own labels (next word) | Reinforcement learning |
| AI workloads | Vision, speech, language, document extraction, prediction, generation | AI workloads |
| Classification, detection, segmentation | One label per image / a box per object / a class per pixel | Task vocabulary |
| When not to use ML | Rules can be written, no data, or every error must be explained | When not to use ML |
| Training data, feature, label | The known examples / an input the model reads (size) / the answer it must predict (price) | The task |
| Parameters | The knobs (weights) training sets | Model parameters |
| Loss | One number for how wrong: MSE for numbers, cross-entropy for classes | Loss |
| MSE, MAE, Huber | Squared: outlier-sensitive / absolute: robust / squared near zero, linear beyond δ | Choosing a loss |
| Gradient descent | Measure the slope, move the knobs downhill | Gradient descent |
| Gradient of MSE | ∂L/∂w = (2/n) Σ error × input; ∂L/∂b = (2/n) Σ error | The gradient |
| Learning rate | Too small crawls; too large overshoots | Learning rate |
| Epoch, batch | One pass over all data / the slice used per step | Epochs and batches |
| Batch, mini-batch, SGD | All examples per step / a few (the default) / one | Batch size |
| Linear regression | Weighted sum of features predicts a number | Linear regression |
| Cleaning | Missing values, outliers, duplicates, wrong types | Data cleaning |
| CSV vs Parquet | Row text / columnar, typed, compressed | Data cleaning |
| One-hot encoding | One 0/1 column per category | Feature engineering |
| Standardise, normalise, log | z-score / 0 to 1 / tame long tails | Feature engineering |
| Train, validation, test | Fit / tune / judge once at the end | Data splits |
| k-fold cross-validation | Every example validates once | Data splits |
| Data leakage | Test or future information reaching training; fit preparation on training folds only | Data leakage |
| Overfitting vs underfitting | Memorised the data / too simple to learn it | Overfitting |
| Regularisation, dropout, early stopping | Fixes for overfitting | Overfitting |
| Learning curve | Training and validation error against training-set size; gap = variance, both high = bias | Learning curves |
| Hyperparameter tuning | Grid, random or Bayesian search over settings | Hyperparameters |
| MAE, RMSE, R² | Mean miss / punishes big misses / share of variance explained | Regression metrics |
| Baseline | Always compare with a simple rule or the majority class | Baselines |
| Decision tree | One question at a time; easy to read, overfits alone | Decision trees |
| Gini impurity | How mixed a node is; the split that lowers it most wins | Decision trees |
| Random forest | Many trees on random samples vote | Random forests |
| Out-of-bag (OOB) score | Each tree is scored on the ~37% of rows its bootstrap sample left out | Why forests work |
| Gradient boosting | Each tree fixes earlier errors (XGBoost, LightGBM) | Gradient boosting |
| eta, max_depth, subsample | XGBoost defaults 0.3, 6, 1; lower eta needs more rounds | Boosting settings |
| k-means | k centres move to the middle of their groups | k-means clustering |
| Elbow, silhouette | Bend in inertia against k / cohesion against separation, from −1 to 1 | Choosing k |
| Choosing | Tables: boosted trees first; images and text: neural networks | Choosing an algorithm |
| Neuron | Weighted sum of inputs plus bias, then an activation | Artificial neuron |
| ReLU, sigmoid, softmax | Bends that let layers learn curves; softmax gives class probabilities | Activation functions |
| Softmax, logits | e^z ÷ Σ e^z turns raw scores (logits) into probabilities that sum to 1 | Softmax and cross-entropy |
| Layers | Input, hidden, output: raw input to features to answer | Layers |
| Forward pass | Input to prediction, layer by layer | Forward pass |
| Backpropagation | The error flows back; each weight gets its share | Backpropagation |
| Vanishing gradient | Slopes below 1 multiplied per layer; fixed by ReLU, normalisation, residual connections | Vanishing gradients |
| CNN | A small filter slides across the image | Convolutional networks |
| CNN, RNN, transformer | Images / sequences (older) / text and most else now | Convolutional networks |
| Conv output size | ⌊(W − K + 2P) ÷ S⌋ + 1; parameters = (K × K × channels in + 1) × filters | Stride and padding |
| GPU | Matrix maths in parallel | GPUs and accelerators |
| Lifecycle | Frame, data, features, train, evaluate, deploy, monitor, repeat | ML lifecycle |
| Problem framing | A business goal becomes one prediction you can measure | Problem framing |
| Real-time, batch, async inference | Pick by who waits and how often requests come | Inference options |
| p95, p99 latency | The time 95% / 99% of requests finish within; targets use these, not the mean | Sizing an endpoint |
| MLOps | Versioned data and models, pipelines, registry, monitoring | MLOps |
| Feature store | Offline store for point-in-time training sets, online store for serving; one definition | Feature stores |
| Shadow, canary, A/B, blue-green | Shift traffic to a new model in steps | Deployment strategies |
| Real-time, serverless, async, batch endpoints | SageMaker inference options | Deployment strategies |
| A/B test | Split live traffic, compare a business metric, test the difference for significance | A/B testing |
| Data drift, concept drift | Inputs change / the right answer changes | Model drift |
| Model monitor | Compare live inputs and outputs to training | Model drift |
| Token | Piece of a word; roughly ¾ of an English word | Tokens |
| Byte-pair encoding (BPE) | Repeatedly merge the most frequent adjacent pair into a new token | Byte-pair encoding |
| Next-token prediction | The only thing an LLM does, again and again | Next-token prediction |
| Embedding | Meaning as a point; close when similar | Embeddings |
| Cosine similarity | a·b ÷ (|a| |b|); equals the dot product for unit-length vectors | Cosine similarity |
| Attention | Each token draws on every other token | Attention |
| Long context cost | Attention grows with the square of the length | Attention |
| Context window | Everything the model can see: prompt, history, answer | Context window |
| Pretraining, fine-tuning, RLHF | Language / instructions / preferences | Training stages |
| Foundation model | One pretrained model adapted to many tasks | Training stages |
| Temperature | Low: focused and repeatable; high: varied | Temperature and sampling |
| Top-p, top-k, max tokens, stop | Other sampling controls | Temperature and sampling |
| BLEU, ROUGE, BERTScore, LLM judge | Generated text by overlap, meaning or a judge | Temperature and sampling |
| Diffusion | Start from noise, remove it step by step | Diffusion models |
| Prompt parts | Task, context, example, format | Prompt anatomy |
| Zero-shot, few-shot, chain of thought | Techniques and when to use each | Prompt anatomy |
| System prompt | Sets rules for the whole chat; messages carry turns | System prompt and messages |
| system / system_instruction | Where each API takes the system prompt; Gemini calls the assistant role "model" | API differences |
| Few-shot | Two or three examples teach the format | Few-shot prompting |
| Step by step | Ask for the reasoning first on multi-step problems | Step-by-step reasoning |
| Structured output | JSON against a schema when code reads the answer | Structured output |
| Validate and retry | json.loads, schema validation, one retry with the error text, then hand off | Validate and retry |
| Prompt failures | Vague task, missing context, conflicting rules | Prompt failures |
| Prompt tests | Version prompts and run a test set on each change | Testing prompts |
| Confidence interval on a score | 45/50 is 79% to 96% at 95% confidence; small sets only show large changes | Test set size |
| RAG | Retrieve relevant passages, put them in the prompt | Why RAG |
| RAG vs long prompt vs fine-tune | Changing facts / small fixed set / behaviour and style | Why RAG |
| Index and query pipelines | Chunk, embed, store once; embed, search, answer each time | The two pipelines |
| Vector store | Bedrock Knowledge Bases, Azure AI Search, pgvector | The two pipelines |
| Chunking | Size and overlap decide what can be found | Chunking |
| Cosine similarity | Nearest chunks in meaning | Vector search |
| ANN index (HNSW) | Approximate search over millions of vectors fast | Vector search |
| Hybrid search, reranker | Keyword plus vector, then a model reorders | Hybrid search and reranking |
| BM25 | Keyword ranking: IDF × saturating term frequency, length-normalised; k1 1.2, b 0.75 | BM25 |
| Grounding | Answer only from the sources, and cite them | Grounding and citations |
| Evaluating RAG | Retrieval (recall@k) and answer (faithfulness) separately | Evaluating RAG |
| Recall@k, MRR, nDCG | Found ÷ existing in top k / mean of 1 ÷ first relevant rank / rank-discounted gain | Retrieval metrics |
| Tool calling | The model asks; your code runs the tool and returns the result | Tool calling |
| Tool definition | Name, description, JSON schema of inputs | Tool calling |
| Agent loop | Reason, act, observe, until the goal is met | The agent loop |
| Managed agents | Bedrock Agents, Azure AI Agent Service, Vertex AI Agent Builder | The agent loop |
| MCP | One standard for apps to use tools and data | Model Context Protocol |
| Host, client, server, transport | The parts of MCP | Model Context Protocol |
| Agent memory | Summarise the context; store facts outside it | Memory |
| Workflow vs agent | Fixed steps when you can; an agent when you must | Workflows and agents |
| Compounding error | A 95%-reliable step gives a 36% chance of 20 clean steps | Why long chains fail |
| Verification step | A check after each agent step; with one retry, 95% a step becomes about 99% | Checks and retries |
| Agent safety | Least privilege; a person approves risky actions | Keeping agents safe |
| Customisation ladder | Prompt, RAG, fine-tune, train from scratch | Ways to customise a model |
| LoRA | Train a tiny add-on instead of every weight | Fine-tuning and LoRA |
| SFT, instruction tuning, preference tuning | Kinds of fine-tuning | Fine-tuning and LoRA |
| Inference cost | Input and output tokens × price per token | Cost of inference |
| Prompt caching | A repeated, identical prompt prefix billed at a reduced rate | Resent history |
| TTFT, tokens per second | Wait before the first token / writing speed; latency = TTFT + output ÷ speed | Latency |
| Quantisation | Fewer bits per weight: smaller, faster, a little less accurate | Quantisation |
| On-demand, batch, provisioned | Pay per token / cheaper, slower / reserved capacity | Quantisation |
| Quantisation scale | Largest |weight| ÷ top integer level; weight is stored as round(weight ÷ scale) | Quantisation by hand |
| Principles | Fairness, reliability and safety, privacy, inclusiveness, transparency, accountability | Responsible AI principles |
| Bias | Compare groups in the data and in the outcomes | Measuring bias |
| DPPL, RD, DAR | Clarify post-training metrics: gap in predicted approvals / in recall / in precision | Bias after training |
| SHAP | How much each feature pushed this one prediction | Explainability |
| Interpretable models | Linear models and small trees explain themselves | Explainability |
| Shapley value | A feature's contribution averaged over all orderings; values sum to prediction − base | Shapley values |
| Hallucination | Fluent and false: ground, cite, check | Hallucination |
| Prompt injection | Instructions hidden in text the model should only read | Prompt injection |
| OWASP LLM Top 10 | Injection, data leaks, excessive agency and the rest | Prompt injection |
| Guardrails | Filters on what goes in and comes out | Guardrails |
| EU AI Act | Unacceptable, high, limited, minimal risk | Governance and regulation |
| ISO/IEC 42001, NIST AI RMF | AI management and risk frameworks | Governance and regulation |
| When the AI Act applies | Placing on, or use in, the EU market | Governance and regulation |
Nothing matches. Try a shorter word.