ML & AI Glossary
Quick reference for terms used across KiranVision courses.
A–C
Activation function — Non-linearity applied after weighted sum (ReLU, sigmoid, tanh).
Backpropagation — Algorithm to compute gradients via chain rule for training neural nets.
Bias-variance tradeoff — Balance between underfitting (high bias) and overfitting (high variance).
Classification — Predicting discrete categories.
Cross-validation — Rotating train/validation splits to estimate generalization.
D–G
Deep learning — ML using neural networks with many layers.
Embedding — Dense vector representation of discrete items (words, users).
Epoch — One full pass through the training dataset.
Feature engineering — Creating informative inputs from raw data.
Gradient descent — Iterative optimization along negative gradient of loss.
H–O
Heuristic — Estimate guiding search (e.g., A*).
Hyperparameter — Settings chosen before training (learning rate, tree depth).
LLM — Large language model trained on text at scale.
MDP — Markov decision process formalism for reinforcement learning.
Overfitting — Model fits training noise; poor test performance.
P–Z
Precision / Recall — Classification metrics for positive class quality and coverage.
Regularization — Penalties (L1/L2, dropout) to reduce overfitting.
Supervised learning — Learning from labeled examples.
Tensor — Multi-dimensional array generalizing scalars, vectors, matrices.
Transformer — Architecture using self-attention for sequences.