Chapter 13: ML Terminology

ML Terminology — all the key words and phrases you keep hearing in Machine Learning (and AI) conversations, books, interviews, or when people talk about ChatGPT, recommendations, or fraud detection.

I’m explaining this like your favorite teacher: slowly, in simple language, with everyday examples (Hyderabad style where possible), analogies, and why each term matters in 2026. No overload — we’ll cover the most important 20–25 fundamental terms that appear everywhere, grouped logically.

Think of these terms as the “vocabulary” of ML. Once you know them, reading papers, code, or job descriptions becomes much easier!

1. Core Building Blocks

• Artificial Intelligence (AI) The big dream: making machines do tasks that normally need human intelligence (seeing, speaking, deciding, learning). Example: Your phone’s face unlock or Google Maps suggesting routes — that’s AI.
• Machine Learning (ML) A subset of AI. Computers learn patterns from data without humans writing every rule. Example: Gmail spam filter learns from millions of emails what “spam” looks like → blocks new spam automatically.
• Deep Learning (DL) Advanced ML using neural networks with many layers (deep). Great for images, voice, text. Example: When Google Photos tags “beach” or “family selfie” automatically — deep learning spots visual patterns.
• Model The “learned program” after training. It takes input → gives prediction/output. Example: A house price predictor model takes size, bedrooms, location → outputs ₹85 lakh.
• Algorithm The step-by-step recipe to build/train a model (e.g., decision tree algorithm, neural network algorithm). Example: “Perceptron algorithm” we talked about earlier — simple rule to adjust weights.

2. Data-Related Terms

• Dataset Collection of examples (rows = samples, columns = features + label). Example: 10,000 Hyderabad flat records (size, bedrooms, price) = dataset.
• Features (Inputs / Independent Variables) The measurable things we feed the model (what we know). Example: In flat price prediction → size_sqft, bedrooms, location_score, age_years.
• Label (Target / Output / Dependent Variable) What we want to predict (the correct answer in supervised learning). Example: price_lakh in flats data, or “spam” / “not spam” in emails.
• Training Data The part of dataset used to teach the model (adjust weights). Example: 80% of your data → train on this.
• Test Data / Test Set Held-out unseen data to check real performance (no cheating!). Example: Last 20% → after training, predict prices → compare to actual prices.
• Validation Set Used during training to tune hyperparameters (not for final test). Example: Decide best learning rate by checking validation accuracy.

3. Learning Types

• Supervised Learning Has labels → learn to predict them. Sub-types: Classification (categories) or Regression (numbers). Example: Spam detection (classification), house price (regression).
• Unsupervised Learning No labels → find hidden patterns/groups. Example: Group Swiggy customers into “frequent biryani lovers” vs “healthy salad buyers” automatically.
• Reinforcement Learning Learn by trial & error + rewards/punishments. Example: Self-driving car gets +reward for safe lane change, -huge for accident.

4. Model Performance & Evaluation Terms

• Accuracy % of correct predictions (simple but misleading if classes imbalanced). Example: 95% accurate spam filter — but if 99% emails are not spam, guessing “not spam” always gets 99%!
• Precision Of all “positive” predictions, how many are actually positive? (low false positives) Example: Cancer detection — high precision means few healthy people wrongly told “cancer”.
• Recall (Sensitivity) Of all actual positives, how many did we catch? (low false negatives) Example: Cancer — high recall means catch most real cancers (don’t miss sick people).
• F1 Score Balance of precision & recall (harmonic mean). Good for imbalanced data. Example: Fraud detection — want both low false alarms & catch most frauds.
• Overfitting Model memorizes training data perfectly but fails on new data (too complex). Example: Perfect score on your 100 Hyderabad flats but predicts badly for new Banjara Hills flat.
• Underfitting Model too simple — poor on both train & test. Example: Using only “size” to predict price — misses location effect.
• Regularization Technique to prevent overfitting (add penalty for large weights). Example: L2 regularization (weight decay) in neural nets.

5. Neural Network & Advanced Terms

• Neural Network / Perceptron / Neuron Basic unit: inputs × weights + bias → activation function. Stack many → deep neural network.
• Epoch One full pass through training data. Example: Train for 50 epochs → model sees all data 50 times.
• Loss Function Measures how wrong predictions are (e.g., Mean Squared Error for regression). Goal: minimize loss during training.
• Gradient Descent How we update weights: follow the “downhill” direction of loss. Example: Adjust weights little by little to reduce error.
• Hyperparameters Settings you choose before training (learning rate, number of layers, epochs). Example: Learning rate = 0.01 → too high → overshoots, too low → slow learning.

6. Quick Everyday 2026 Examples Table

Final Teacher Summary (2026 Advice)

ML terminology is like learning Telugu slang — at first confusing, but once you know 20–30 key words, everything connects! Most interviews/jobs ask about: supervised vs unsupervised, overfitting/underfitting, precision/recall, features/labels, train/test split.

Start using them when you talk about apps:

• “Gmail uses supervised ML classification to detect spam patterns.”
• “My model overfit because accuracy 99% on train but only 80% on test.”

Got the vocabulary now? 🔥

Questions?

• Want deeper explanation of any term (e.g., bias-variance tradeoff)?
• Examples in code (like accuracy vs F1)?
• More advanced terms (like transformers, fine-tuning, RAG)?

Just tell me — next class is ready! 🚀