Learning from examples, not instructions
Traditional programming is explicit: a developer writes the rules, and the computer follows them. Machine learning inverts this. You supply examples and desired outcomes, and the system derives the rules itself by adjusting millions of internal parameters until its predictions line up with reality.
This is why AI can recognise a cat in a photograph — a task nobody has ever successfully written explicit rules for — while still failing at things that seem trivially easy to a person. The system has no understanding of cats. It has a very finely tuned statistical association between certain visual patterns and a label.
A model does not know anything. It has learned which patterns tend to accompany which answers. That distinction explains almost every surprising AI failure.
The parts worth understanding
- Training data — models inherit whatever is in their data, including its gaps and biases. Data quality is not a footnote; it largely determines behaviour.
- Neural networks — layers of simple mathematical units that, stacked deep enough, approximate remarkably complex functions.
- Loss and optimisation — training is a search: measure how wrong the model is, adjust slightly in the direction that reduces the error, repeat millions of times.
- Generalisation — the real test is not performance on data it has seen, but on data it has not. A model that memorises has learned nothing useful.
- Limits — models can be confidently wrong, and they do not signal uncertainty the way a person would.
What students explore with us
- Training a small model from scratch and watching its accuracy climb, plateau, and sometimes get worse
- Deliberately feeding a model biased data to see how faithfully it reproduces that bias
- Working through what a neural network layer actually computes, using numbers small enough to follow by hand
- Testing where models break — adversarial inputs, edge cases, confident errors
- Debating the real questions: attribution, automation, misinformation, and who is accountable when a model is wrong
Where this leads
AI has become infrastructure rather than a speciality, which means the fundamentals now matter across medicine, finance, climate modelling, robotics and research. Students who understand the mechanism — rather than only the interface — are the ones who can apply it responsibly and spot when it is being misapplied. Given how much of the global AI workforce is based in or trained in India, this is a field where early grounding compounds quickly.