An objective function turns prediction quality into a number that can guide learning. For a neural network, the loss measures mismatch between predictions and observed data: lower loss means a better fit. section 10.1
A neural network is a function f(x; w), where w contains its weights and biases. Training searches for parameter values that minimize the loss. A common choice is average squared error: calculate each prediction error, square it, and average the results. Gradient descent then moves the parameters toward lower loss. section 10.1
Fitting noisy sine data
A network trained on noisy observations generated from a sine function can achieve a low average squared error and interpolate the observed points well. The loss helps fit the data, but the resulting predictions do not automatically represent observation noise or uncertainty in sparse regions. section 10.1
The same average squared error is also used for linear regression. In Bayesian reasoning, loss can instead describe the consequences of competing decisions, such as deploying or not deploying a website. section 10.1 section 11.0









































