IB Computer Science Topic A4.1.1: Types of Machine Learning and Their Applications in the Real World Notes

Machine Learning: Patterns Learned From Data Replace Rules Written by Hand

1. The program builds a model from training data, then uses that model to detect patterns, assign categories or predict outcomes for inputs it has never seen.
2. Deep learning is a subset of machine learning, so the three terms nest: artificial intelligence contains machine learning, which contains deep learning.
3. The five approaches on this syllabus differ in two things: what the training data looks like, and what feedback the model receives while it learns.

Deep learning and transfer learning describe how a model is built and where its training starts from, so either can sit on top of any of the first three approaches.

Supervised Learning: Labelled Examples Give the Model an Answer to Copy

1. The model adjusts itself to reduce the difference between its prediction and the label, so the label is the teaching signal.
2. Classification predicts a discrete category.
   1. Examples include deciding whether an email is spam or not spam, or which of ten digits a handwritten image shows.
3. Regression predicts a continuous numerical value.
   1. Examples include the selling price of a house, or tomorrow's electricity demand in megawatts.
4. The characteristic cost of supervised learning is labelling, because a person has to supply the correct answer for every training example.

Unsupervised Learning: Structure Found With No Labels at All

1. Clustering groups records that are similar in their features.
   1. A streaming service can cluster listeners by what they play and when, then treat each cluster as a segment without deciding the segments in advance.
2. Association rule learning finds attributes that occur together across many records.
   1. It reports rules of the form if a record contains X, it often also contains Y.
3. The characteristic difficulty is evaluation, because with no labels there is nothing to check the output against and the result has to be judged on whether it is useful.

Reinforcement Learning: Consequences Replace Worked Examples

1. The agent is never shown a correct action, so it has to discover which actions pay off by trying them.
2. Its goal is to maximise cumulative reward across a whole sequence of actions, not the reward from the next action alone.
3. This suits problems where the outcome depends on a series of decisions and the payoff arrives late, such as a robot crossing an unfamiliar room or a program playing chess.

Deep Learning: Layers That Work Out Their Own Features

1. Each layer transforms the output of the layer before it, so early layers capture simple features such as edges, and later layers combine them into complex ones such as faces.
2. That layering removes the need for a programmer to decide which features matter, because the network learns the features from the data.
3. The cost is data and processing power, which is why A4.1.2 covers GPUs, TPUs and cloud platforms.

Transfer Learning: Reusing a Model That Has Already Been Trained

1. The pre-trained model already encodes general features, so only the later layers need adjusting to the new task.
2. Fine-tuning in this way needs far less labelled data and far less training time than training a model from scratch.
3. It works only when the source and target tasks are related, because the reused features have to be relevant to the new data.

Six Real-World Applications and the Approach Each One Needs

1. Market Basket Analysis
   1. A retailer mines transaction records to find products that are bought together, then uses the result for shelf layout and cross-selling.
   2. This is unsupervised and uses association rule learning, because no transaction is labelled with the answer.
2. Medical Imaging Diagnostics
   1. A model sorts X-rays, CT scans or MRI scans into diagnostic categories for a radiologist to confirm.
   2. It is supervised and normally deep, and hospitals often add transfer learning because labelled scans are scarce.
3. Natural Language Processing
   1. Systems translate text, answer questions and extract meaning from written and spoken language.
   2. This relies on deep learning over very large text collections, with transfer learning adapting one general language model to many specific tasks.
4. Object Detection and Classification
   1. A model locates and names objects inside an image, which is what lets an autonomous vehicle separate a pedestrian from a lamp post.
   2. It is supervised and deep, trained on images in which every object has been outlined and labelled by hand.
5. Robotics Navigation
   1. A robot learns a route through a space by acting and being rewarded for progress, so this is reinforcement learning.
   2. No labelled data set of correct movements exists beforehand, and the robot generates its own data by moving.
6. Sentiment Analysis
   1. A model reads reviews or messages and classifies the writer's attitude as positive, negative or neutral.
   2. It is supervised classification, and transfer learning adapts a general language model to the vocabulary of one industry.

Choosing the Approach From the Data You Have

1. Read the stem for the data, not for the industry, because one industry uses different approaches for different tasks.
2. A hospital classifying scans against confirmed diagnoses is supervised, while the same hospital grouping patients by unlabelled test results is unsupervised.
3. If the system has to act before any data exists, the approach is reinforcement learning whatever the sector.