Peek LLM
/
Backpropagation
Studio · Computed tiny MLP gradients
Fig. 1
The forward pass has finished; probabilities are ready.
A computed four by three by three neural network with a loss node
The forward pass ends in three token probabilities and a cross-entropy loss. The backward pass sends error signals from the loss back through every wire to produce a gradient for each weight. Select a node or wire to inspect its chain-rule calculation.
Forward.
The same computed teaching network as Forward Pass Studio ends in a loss.
Select any node or wire to inspect it.