How to read this view
Brightness shows activation. Hidden layers use a fixed exponential scale controlled by exposure. Predictions use absolute probability. Original-model negative logits have dark fill. ResNet and ViT signed values use magnitude brightness; inspect a cell for its sign.
Overview samples up to six evenly spaced channels per convolution. Focus reveals every channel. All visible feature maps update together from the latest inference. Overview lines show sampled receptive fields and four strongest weights per dense neuron; brightness follows absolute input contribution.
Dense lines show the top 12 terms for the selected neuron. Solid = positive weight, dashed = negative. Softmax is a shared normalization, not a learned connection.
ResNet bypass lines are sampled identity shortcuts. ViT shows 16 patches, token feature strips and three attention matrices. Select a head and query token to see its strongest patch relationships. Attention is a mixing weight, not feature importance. Its brightness uses a fixed square-root scale; signed token features use magnitude. Intermediate Transformer operations remain available in Layer.