Scalar Computation Graph
Key concept: Backpropagation applies the chain rule locally and adds gradients from shared paths.
Build a graph of + − × ÷ nodes. Values move forward; gradients move back. In Chain rule mode, select a node to see every path and their summed products. In Local gradient mode, press Step to receive, compute, and pass each gradient.
What to look for: the share preset shows that an input’s gradient is complete only after both uses contribute.
Open the demo in a new tab for more room.