Learn AI / Term Explorer /How Models Improve
How Models Improve
The machinery of getting less wrong: measuring error, working out which settings caused it, and nudging them.
Starter
Intermediate
Backpropagation
A method that propagates derivatives backwards through a model to compute how changes to each weight would affect the loss.
Convergence
Convergence means training updates are settling towards a stable, lower-loss solution instead of continuing to wander or blow up.
Gradient
A gradient tells you how much a model’s error would change if each setting were nudged slightly up or down.
Gradient descent
A method for reducing a model’s loss by repeatedly adjusting its internal settings in the direction opposite the loss gradient.
Hyperparameter
A hyperparameter is a training setting chosen before or during learning that shapes how a model is fitted, rather than a value the model learns from data.
Learning rate
A training setting that controls how far model parameters move after each error-based update.
Loss function
A loss function turns a model’s output and the correct target into a single number that shows how wrong the model was for that example or batch.
Optimisation
The process of adjusting a model’s internal settings to reduce loss, plus the algorithms and rules used to choose each adjustment.
Deeper
Adam optimiser
A training method that adjusts each model setting using both recent gradient direction and the recent size of past gradients.
Chain rule
A calculus rule for tracing how a small change in one setting affects the final error through a chain of intermediate calculations.
Dropout
A training method that randomly switches off some units during updates to reduce reliance on particular pathways and potentially improve generalisation.
Local and global minimum
A local minimum is the lowest loss in a nearby region, while a global minimum is the lowest loss anywhere in the whole set of possible model settings.
Momentum
A training method that carries part of the previous update into the next one so parameter changes keep moving in useful directions and wobble less.
Regularisation
Regularisation changes a model's training objective or procedure to reduce overfitting and improve performance beyond the training data.
Vanishing and exploding gradients
Training can stall or blow up when error signals shrink toward zero across layers or time steps, or grow so large that updates become unstable.