Backpropagation, step by step
Forward pass, loss, gradients, update - shown one calculation at a time on a real network you can change.
The idea in four moves
1. Forward: the input flows through each layer - weighted sums, then activations - to a prediction.
2. Loss: the prediction is compared with the right answer.
3. Backward: the chain rule sends the error back, layer by layer, giving every weight a gradient - how much the loss would change if that weight moved.
4. Update: the optimiser (plain gradient descent, momentum, Adam...) nudges every weight against its gradient. Repeat.
See it, not just read it
NeuroScope pauses after each of these moves. Pick an example, follow it forward, see the loss, and watch the gradients appear on every connection - with the numbers that produced them.
Change the learning rate, the batch size or the activation and see what each one does to the gradients and to learning.
Try it in a minute
- Press Try it now.
- On the Train page press Start, then Next step - forward first, then backward.
- Click any layer to open its exact calculations.
Questions
Is this the real algorithm or a simplified animation?
The real one: every gradient comes from the same calculation that trains the network, and the page checks it against the layer's actual output.
Can I try different optimisers?
Yes - SGD, mini-batch, full batch, momentum, Adam and more, with schedules and weight decay.
Does it work on my phone?
Yes, small networks run in any modern browser; bigger image models are happier on a laptop.