Network visualization with real weights
Every node in the diagram corresponds to an actual neuron / feature of the network, and its color and values come straight from the trained checkpoint. Hover over the nodes — click to pin a selection and see the real weights, norms and strongest connections.
The first node of each layer — actual input and output
We ran the network on a sample game state (step 57, my goose of length 5) and captured the values in every layer. The key answer to the question “why does a 3×3 filter produce a 7×11 output?”: the filter really has only 3×3 weights per channel, but it slides over all 7 × 11 = 77 positions of the map (thanks to circular padding, across the edges too) and computes one number at each position. So a single filter returns not a single number but a whole 7×11 map.