AI / ML / Python Glossary

Every key term from this book — the vocabulary of a working ML engineer.

Accuracy

Fraction of correct predictions over all predictions. Misleading on imbalanced datasets.

Activation function

Non-linear transform between neural net layers — sigmoid, ReLU, tanh.

Backpropagation

Algorithm to compute gradients for every weight in a network. The ‘learning’ step.

Bias (fairness)

Unfair behaviour of an AI, usually from biased training data.

Classification

Predicting a category — spam/not, cat/dog/elephant.

Confusion matrix

A grid showing the TP, FP, FN, TN of a classifier.

CSV

Comma-Separated Values file format. Most ML datasets ship as CSVs.

DataFrame

A pandas table — rows are records, columns are features.

Decision tree

Model that learns yes/no questions. Explainable. Used in Niramai.

Deep learning

Neural networks with many layers (typically 5+).

Deploy

Make a trained model available to users via web app, API, or mobile.

DPDP Act 2023

India’s law on personal data — consent, deletion, fines up to Rs 250 crore.

Epoch

One full pass over the training dataset.

F1 score

Harmonic mean of precision and recall. Use when both matter equally.

Feature

An input column in your dataset. Age, salary, pixel value — all features.

Hugging Face

The GitHub of AI. Hosts 30 lakh+ models. Free Spaces for deployment.

Joblib

Python library to save/load trained sklearn models. Industry standard.

Linear regression

Predict a number from inputs using a y = m×x + b style equation.

MNIST

Handwritten digit dataset. 60K images. The standard deep-learning benchmark.

Model card

Document about a model — purpose, data, performance, limitations, contact.

Neural network

Stack of perceptrons in layers. Learns complex patterns.

Overfitting

Model memorises training data but fails on new data.

pandas

Python library for data manipulation. DataFrames live here.

Precision

TP / (TP + FP). When you say ‘positive’, how often are you right?

PyTorch

Deep learning framework. Used by ChatGPT, Tesla, Sarvam, ISRO.

Random Forest

Many decision trees voting together. Strong, explainable, popular.

Recall

TP / (TP + FN). Of all real positives, how many did you catch?

ReLU

Activation function: max(0, x). Standard in deep learning today.

R²

Goodness-of-fit score for regression. 1.0 = perfect, 0 = no better than the mean.

scikit-learn

Python’s classic ML library. Used in most Indian ML jobs.

Sigmoid

Activation function squashing any number into the 0-1 range.

Streamlit

Python library that turns ML scripts into web apps. Free hosting.

Tensor

Multi-dimensional array. PyTorch’s basic data type.

Test set

Held-out data the model has NEVER seen. Used only for the final measure.

Train/test split

Splitting data — usually 80% train, 20% test, with a fixed seed.

Transformer

Modern neural net architecture. Powers ChatGPT, Krutrim, BERT.

Underfitting

Model is too simple — fails to learn patterns even in training data.