Learning in ML is the process by which a model improves at a task by adjusting its internal parameters based on examples. Show a model thousands of cat photos labeled 'cat' and 'not cat', and it learns to recognize cats. The mechanism behind this — gradient descent — is the engine of nearly all modern AI.

When we say a machine learning model 'learns', we mean something specific. The model has internal parameters (weights), and those weights start out random. Show the model a training example along with the correct answer, and the model produces its current prediction. The difference between the prediction and the correct answer is the error, calculated by a loss function. Then comes the key step: gradient descent. The model computes how much each weight contributed to the error and adjusts each weight slightly in the direction that would reduce the error. Repeat this billions of times across millions of examples, and the weights converge to values that make the model's predictions accurate. Different learning paradigms approach this differently. Supervised learning uses labeled examples — the canonical setup. Unsupervised learning finds patterns in unlabeled data. Reinforcement learning learns from reward signals received after taking actions. Self-supervised learning generates its own labels from unlabeled data, which is how language models like GPT learn from raw text. Transfer learning starts with a model pretrained on one task and fine-tunes it for another, dramatically reducing the data needed for specialized models. The remarkable thing about modern ML is that this fundamentally simple mechanism — adjust weights to reduce error — produces systems that can recognize speech, generate images, play games at superhuman levels, and write coherent essays. The learning process itself is surprisingly elegant given the complexity of what it produces.

What is Learning in Machine Learning?