Perspectives on Machine Learning (DP IB Theory of Knowledge): Revision Note

Perspectives on machine learning

• Machine learning is a type of AI where a computer system learns patterns from data in order to make predictions or decisions, improving its performance through training rather than being explicitly programmed with fixed rules

Definitions of ‘knowing’ applied to technology

• In ToK, “knowing” usually implies more than producing correct outputs; it can include understanding and the ability to explain why a certain decision has been reached

• Machines can be described as “knowing” in a limited sense if they can reliably recognise patterns and produce useful predictions from data

• Whether a machine “knows” depends on the standard being used, e.g.:

  • performance-based: does it get the right answers consistently?

  • justification-based: can it give reasons that make sense to humans?

  • understanding-based: can it apply ideas flexibly in new contexts?

AI learning vs human understanding

• Machine learning models “learn” by adjusting the model to fit patterns in training data; they optimise for performance, not meaning or intention

• Human understanding often involves causal explanations and awareness of context, so humans can explain why something might or might not be true

• AI can be strong at correlation-based tasks (finding patterns) but weak at causal reasoning unless it is specifically designed and tested for it

  • E.g. an AI tool spots that certain symptoms often occur together, but a doctor explains the underlying cause and rules out alternatives

• Humans can revise beliefs using broader background knowledge and values, while AI is constrained by the data and objective on which it was trained

Bias embedded in training data

• Training data is the dataset a machine learning system learns from, usually made up of many examples (inputs) with outcomes or labels the system is trying to predict

  • E.g. thousands of photos labelled with the person’s name, used to train an AI to recognise the person

• Training data is determined by human choices, such as how data is labelled and what is included, so patterns learned by AI can reproduce existing biases, e.g. due to:

  • unrepresentative data in which some groups are under sampled

  • biased labels

  • historical patterns that reflect unfair practices

• Biased training data can lead to systematically different error rates across groups, harming the reliability of knowledge claims

  • E.g. a facial recognition system performs worse on darker skin tones because the training set contained fewer examples

• Bias may be hidden because outputs look objective, so careful evaluation is needed