The Computational Magic of the Ventral Stream: Towards a Theory

I conjecture that the sample complexity of object recognition is mostly due to geometric image transformations and that a main goal of the ventral stream – V1, V2, V4 and IT – is to learn-and-discount image transformations. The most surprising implication of the theory emerging from these assumptions is that the computational goals and detailed properties of cells in the ventral stream follow from symmetry properties of the visual world through a process of unsupervised correlational learning.

From the assumption of a hierarchy of areas with receptive fields of increasing size the theory predicts that the size of the receptive fields determines which transformations are learned during development and then factored out during normal processing; that the transformation represented in each area determines the tuning of the neurons in the aerea, independently of the statistics of natural images; and that class-specific transformations are learned and represented at the top of the ventral stream hierarchy.

Some of the main predictions of this theory-in-fieri are:
1. the type of transformation that are learned from visual experience depend on the size (measured in terms of wavelength) and thus on the area (layer in the models) – assuming that the aperture size increases with layers;
2. the mix of transformations learned determine the properties of the receptive fields – oriented bars in V1+V2, radial and spiral patterns in V4 up to class specific tuning in AIT (eg face tuned cells);
3. invariance to small translations in V1 may underly stability of visual perception
4. class-specific modules – such as faces, places and possibly body areas – should exist in IT to process images of object classes