PROJECT FILM / ZONGO MAQUTU

Inside a neuron: contributions, gradients and weight maps

Name: Inside a neuron: contributions, gradients and weight maps
Uploaded: 2026-10-05T01:22:53.911Z
Duration: PT24.88S
Description: Inspect hidden neuron 2 in the fourth recorded MNIST batch, then examine the first-layer signed input-weight maps. Recorded application values. Sample neuron derivatives belong to the displayed digit; weight gradients belong to the mean batch loss. Scrubbing a committed step does not retrain the model. Weight maps are parameters, not training images.

Inspect hidden neuron 2 in the fourth recorded MNIST batch, then examine the first-layer signed input-weight maps.

Inside hidden neuron 2 · Forward → Backward → Update → Compare · signed weight maps, not input images. · Silent demonstration.

What the film shows

Weighted contributions and bias reconstruct the neuron’s preactivation.
Forward, Backward, Update, and Compare expose the same recorded step at different phases.
The Weights panel reshapes learned parameters into 28 × 28 tiles.

Behind the demonstration

Recorded application values. Sample neuron derivatives belong to the displayed digit; weight gradients belong to the mean batch loss. Scrubbing a committed step does not retrain the model. Weight maps are parameters, not training images.

Built and documented by Zongo Maqutu. Read the Field Note for the implementation, architecture and limitations.

Read Field Note Launch Project