Backprop Detective

How Backprop Detective Works

A small feedforward net was shipped with the right blueprint, but production silently flipped exactly one weight. Now it hallucinates: the output is wrong. You are internal affairs. Trace the blame backward through the layers and accuse the single guilty edge.

1. You see the inputs, the intended weights (every edge is labeled), the Expected output and the wrong Got output.
2. Each hidden neuron's value should be the sum of (incoming weight × input). Do the multiply-add in your head.
3. Spend a Probe to reveal a hidden neuron's actual activation. If it disagrees with your math, the fault is one of that neuron's input weights.
4. If every hidden neuron checks out but the output is still wrong, the traitor is a hidden→output weight.
5. Switch to Accuse, tap the guilty edge. Fewer probes used = higher score. Each solved case grows the network.

Pinning the Exact Weight

Say neuron H2 should be 14 but probes back 20 — it's off by 6. The only incoming input worth 3 means the corrupted weight is the one multiplying that input (6 / 3 = a clean +2 drift). The other inputs don't divide evenly, so they're innocent. That clean-division test is your fingerprint kit.