Explore how activations shape neuron outputs and how loss functions measure prediction error โ then see how they chain together in a forward-backward pass.
Toggle functions on/off. Drag the input x slider to see how each activation transforms the input and what gradient it produces. Watch for vanishing gradients (flat zones) and dead neurons (zero output regions).
Drag predicted value to see how each loss function penalizes the error. Notice how cross-entropy heavily penalizes confident wrong predictions, while MSE grows quadratically.
See a complete forward-backward pass: raw logit z โ activation a = f(z) โ loss L(a, y) โ gradient โL/โz = โL/โa ยท f'(z). Observe how activation choice affects gradient flow.