A4.1.1 — Real-World Applications

Overview

For each application you need to know: what it does, which learning paradigms apply, and why each paradigm is appropriate. The exam regularly asks you to justify your choice.

SL Supervised UL Unsupervised RL Reinforcement DL Deep Learning TL Transfer Learning

Market Basket Analysis

Identifies associations between items purchased together. Algorithms (Apriori, FP-Growth) find which products are likely to be bought together. Used for cross-selling in retail.

Why unsupervised? There are no predefined labels or target outputs. The goal is to discover purchase patterns from raw transaction data without prior knowledge of relationships. Association rules (X implies Y) are derived purely from the data structure.

Medical Imaging Diagnostics

ML models trained on MRIs, X-rays, and CT scans to identify and diagnose medical conditions.

SL DL TL

Supervised: Each image is labelled with a diagnosis. The model learns to map image features to diagnostic labels.
Deep learning (CNN): CNNs automatically detect intricate patterns in pixel data without manual feature extraction. They learn hierarchies from simple edges to complex pathological features.
Transfer learning: Labelled medical datasets are scarce due to privacy and expert annotation requirements. A model pre-trained on general image tasks is fine-tuned on medical images, leveraging learned visual features (edges, textures) that transfer across domains.

Natural Language Processing (NLP)

Processing and understanding human language. Tasks include translation, sentiment analysis, and chatbots. Key models: BERT, GPT (transformer-based).

SL DL TL

Supervised: Tasks like sentiment analysis and translation use labelled data where text is paired with target outputs.
Deep learning (Transformers, RNNs): Language is sequential and contextual. Transformers use self-attention to weigh the relevance of all words simultaneously, regardless of position. RNNs and LSTMs handle long-range dependencies in sequential text.
Transfer learning: Training language models from scratch requires enormous data. BERT is pre-trained on massive text corpora then fine-tuned on specific tasks (e.g. sentiment analysis for reviews) with much less labelled data.

Object Detection and Classification

Identifying and categorising objects within images, including their locations (bounding boxes). Used in autonomous vehicles, security cameras, and quality control.

SL DL

Supervised: Training images are annotated with labels and bounding boxes. The model learns to predict both the presence and position of objects in new images.
Deep learning (CNN, YOLO, R-CNN): CNNs recognise spatial hierarchies in images. Architectures like YOLO (You Only Look Once) process the whole image at once for real-time detection. R-CNN identifies candidate regions before classification.

Robotics Navigation

Robots learning to navigate and manipulate their environment. They must avoid obstacles, plan paths, and respond to changing conditions.

RL SL

Reinforcement learning: Robots interact with unpredictable environments. They learn optimal behaviours through trial and error, receiving rewards for correct actions and penalties for mistakes. Ideal for sequential decision-making across changing states.
Supervised: Used for specific sub-tasks like obstacle detection or terrain classification where sensor readings are paired with correct labels. Allows immediate, accurate response to environmental cues.

Sentiment Analysis

Analysing text (reviews, social media) to determine expressed sentiment: positive, negative, or neutral.

SL DL TL

Supervised: Text is annotated with sentiment labels. The model learns to correlate word choice and sentence structure with sentiment.
Deep learning (LSTM, Transformer): LSTMs handle long-range dependencies in text. Transformers use self-attention to capture complex contextual relationships between words across the entire input.
Transfer learning: Domain-specific labelled data is often limited. A model pre-trained on general language tasks is fine-tuned on domain-specific text (e.g. product reviews), significantly reducing data requirements.

Summary Table

Application	Paradigms	One-line reason
Market basket analysis	UL	No labels. Discover purchase patterns from raw data.
Medical imaging	SL, DL, TL	Labelled images, CNNs for pixels, transfer because data is scarce.
NLP	SL, DL, TL	Labelled text tasks, transformers for context, transfer for efficiency.
Object detection	SL, DL	Annotated bounding boxes, CNNs for spatial feature detection.
Robotics navigation	RL, SL	RL for environment interaction, SL for specific sensor classification.
Sentiment analysis	SL, DL, TL	Labelled text, LSTMs or transformers for language, transfer for limited data.

Ninja Note: Always justify each paradigm separately. Saying "supervised learning is used because the data is labelled" is not enough. Say what the labels are, what the model learns to predict, and why that matters for this specific application.