Vision: Learning the Filters

Part I ended with a prescription: make the template local, and slide it. We begin with filters whose numbers are fixed so that locality, sharing, and translation equivariance are visible before learning enters. A convolutional network keeps that spatial structure built in and makes the detector weights learnable. Channels let detectors combine, depth expands each receptive field, and pooling buys some shift tolerance by discarding exact position. Normalization and residual paths help deeper versions train, while transfer asks which learned features deserve to be reused. After this Part, the autoencoder interlude compresses those features into one code and exposes the next boundary: a one-shot fixed-width code is not a variable-length process.