This segment details the use of categorical cross-entropy loss in conjunction with softmax activations. It explains how this combination addresses the limitations of sigmoid, providing effective gradient descent updates. The explanation includes a breakdown of how the loss function interacts with one-hot encoded targets and how it leads to non-zero deltas for all inputs, even when targeting zero outputs, ensuring effective training. For multi-class classification, sigmoid output neurons are problematic. Softmax activations, combined with categorical cross-entropy loss, provide a superior solution. Softmax normalizes outputs to sum to 1, improving interpretability. Categorical cross-entropy avoids the vanishing gradient problem inherent in sigmoids, enabling efficient training. This section introduces the softmax activation function as an alternative to sigmoid for multi-class classification. It describes how softmax computes activations across the entire layer, normalizing them to sum to one, magnifying differences between inputs, and pushing the largest input towards one while others approach zero, resulting in more interpretable outputs. This segment explains the limitations of using sigmoid neurons in the output layer for multi-class classification. It highlights how sigmoid outputs can be difficult to interpret due to multiple large values, and how the vanishing gradient problem hinders effective training when using one-hot encoded target vectors, leading to slow convergence and difficulty correcting confidently wrong predictions.