IB Digital Society 3.6B Types and Uses of Machine Learning Notes

3.6B Types and Uses of Machine Learning

What machine learning is

1. Machine learning powers almost all of today's AI. The programmer does not write every rule. The system is shown large data, works out the patterns itself, and uses them to make predictions.
2. In traditional programming a human writes the rules and the computer applies them. Machine learning works differently. Give it thousands of emails already marked spam or not, and it works out the rules itself.
3. This suits tasks too messy to write rules for, like telling a cat from a dog in a photo. The catch is the data: a model is only as good, and only as unbiased, as what it learned from.

Supervised learning

1. Supervised learning is the most common type: the training data is labelled (each example tagged with the correct answer), and the model learns to map input to label so it can predict labels for new inputs.
   1. Mechanism is a loop of guess and correct: predict, compare to the true label, measure the error, nudge internal settings, repeat millions of times. Drives spam filters, medical-image classifiers, and credit scoring.
2. It is only as good as its labels. If training photos were one breed, or past credit decisions were biased, the model quietly repeats that pattern while looking objective, so data bias begins with the human label-makers.

Unsupervised learning

1. In unsupervised learning the model is given unlabelled data and finds structure on its own, discovering patterns or groupings not specified in advance.
   1. The classic task is clustering: sorting data into groups of similar items. A shop can cluster customers by buying habits and discover a group it never knew existed, all without defining categories first.
2. Risk: the groups have no built-in meaning and may track race, income, or postcode. If a company then treats clusters differently, it can discriminate without naming the trait. Whether such a grouping is useful or unfair depends on how the company chooses to use it.

Reinforcement learning

1. In reinforcement learning there is no fixed dataset of answers. An agent acts in an environment and learns by trial and error: good outcomes earn a reward, and it adjusts to maximise reward.
   1. DeepMind systems learned to play Atari games and Go purely by trial and reward, beating the best humans with moves no coach had taught them. Designers set the reward, and the system works out the strategy.
2. The choice of reward is the heart of the design. An agent optimises what it is scored on, not what designers meant: a boat-racing agent once looped to collect points instead of finishing, a case of unintended consequences (reward hacking).

Deep learning and neural networks

1. Deep learning is ML built on neural networks with many layers. Data passes through layers of connected units, and each layer detects finer patterns than the one before.
   1. Deep means many layers: for a face, early layers detect edges, middle layers eyes and noses, later layers the whole face, learned from raw data. It removed the need for hand-designed features and powers speech recognition, translation, and image generation.
2. The cost is scale and opacity. It needs enormous data and computing power, which concentrates capability in a few large organisations. Its many layers also make it a black box: highly accurate, yet hard to explain even for the people who built it.

The uses of machine learning

1. Pattern recognition (finding regularities in data) is the umbrella: facial recognition identifies a person, speech recognition turns audio into text, and natural language processing handles translation, chatbots, and search.
2. The same capability is contested. Facial recognition can find a missing person, but it also enables mass surveillance and misidentifies people. Speech systems widen access, yet mishear some accents far more often. The benefits and harms fall on different groups of people.
3. Most harm here traces back to the data, not to malice. Gaps in the training data show up as gaps in performance, and they fall on the groups already least represented.