Machine learning fundamentals

The layers again, precisely

AI is the umbrella: intelligent systems doing perception, reasoning, learning, problem solving, decisions. Rule-based expert systems (1970s MYCIN) are AI but not ML.

ML is machines learning from data to predict, with no explicitly programmed rules. Regression and classification live here.

Deep learning is neural networks: nodes in layers, connections adjusting as data flows through. It handles complex patterns (computer vision, NLP) and needs lots of data and GPUs. The intuition: each layer learns a pattern (edges, then curves, then digits).

GenAI is the deep learning subset built on foundation models. Transformers process whole sentences at once with self-attention, weighting which words matter (GPT = Generative Pretrained Transformer; BERT reads both directions).

Multimodal models mix input and output types: text + image + audio in, video out.

Glossary the exam likes

GPT (text and code generation), BERT (bidirectional text), RNN (sequential data: time series, speech), ResNet (a deep CNN for images and faces), SVM (classification and regression), WaveNet (audio synthesis), GAN (synthetic data generation, data augmentation), XGBoost (gradient boosting).

Training data

Garbage in, garbage out: data is the most critical stage.

Labeled data (features plus known outputs) feeds supervised learning; unlabeled feeds unsupervised. Structured means tabular rows and columns or time series; unstructured means free text and images.

Supervised learning

Learns a mapping from labeled examples to predict outputs for new inputs.

Regression predicts a continuous number (house price, stock, temperature). Classification predicts a category: binary (spam or not), multiclass (mammal, bird, reptile), multi-label (a movie tagged action and comedy). k-NN is the flagship algorithm.

The dataset split: about 80% training, 10% validation (hyperparameter tuning), 10% test (final evaluation).

Feature engineering uses domain knowledge to extract, select, and transform features (birthdate becomes age, price per square foot). It’s a big lever for structured data; for text and images it’s TF-IDF, embeddings, CNN features.

Unsupervised learning

Find structure in unlabeled data; humans name the groups afterward.

Clustering (K-means) does customer segmentation. Association rule learning (Apriori) does market basket analysis: bread goes with butter. Anomaly detection (Isolation Forest) does fraud flagging.

The in-betweens

Semi-supervised: a small labeled set plus a large unlabeled set; the model pseudo-labels the rest, then retrains.

Self-supervised: the model invents pretext tasks from unlabeled data (predict the masked word, predict the next part) to learn representations, then tackles downstream tasks. This is how BERT and GPT got made.

Reinforcement learning got its own post: RL and RLHF.