1. What does a hidden layer build?
See each hidden unit’s activation map next to the final decision region. Change boundaries and output weights and observe exactly how a nonlinear classifier is assembled.
Open experimentHow to use these: before changing a control, make a prediction. After the picture changes, explain the change in terms of the model equation, not only in terms of “the network learned.”
2. Forward-pass microscope
Select a hidden unit and see its preactivation broken into individual weighted contributions, then watch the activation and logits update as inputs move.
Open experiment3. Backprop microscope
Select any edge and decompose its gradient into upstream gradient × local derivative. Preview the exact SGD update and its effect on the loss.
Open experiment4. Why gradients vanish or die
Build a deeper scalar network and watch forward activations and backward gradients layer by layer for sigmoid, tanh, and ReLU.
Open experiment5. Training dynamics laboratory
Compare capacity and optimization on the same fixed dataset and initialization. Watch the decision surface, loss curve, hidden boundaries, and train/test accuracy together.
Open experiment