Learning to discount transformations as the computational goal of visual cortex

It has been long recognized that a key obstacle to achiev-inghuman-levelobjectrecognitionperformanceis the prob-lem of invariance [10]. The human visual system excelsat factoring out the image transformations that distort ob-ject appearance under natural conditions. Models with acortex-inspired architecture such as HMAX [9,13] as wellas nonbiological convolutional neural networks [5] are in-variant to translation (and in some cases scaling) by virtueof their wiring. The transformations to which this approachhas been applied so far are generic transformations; a sin-gle example image of any object contains all the informa-tion needed to synthesize a new image of the tranformedobject [15]. In a setting in which transformation invariancemust be learned from visual experience (such as for a new-bornhumanbaby),we haveshown that it is possible to learnfrom little visual experince how to be invariant to the trans-lation of any object [7]. The same argument applies to allgeneric transformations.Generic transformations can be “factored out” in recog-nition tasks (see figure 1) and this is key to goodrecognitionperformance. This is the reason underlying recent observa-tions that random features often perform well on computervision tasks [4,6,11,12].For simplicity consider a specific example: HMAX. Inan architecturesuch as HMAX, if an input image is encodedin terms of similarity to a set of templates (typically via adot product operation) and if the encoding is made invari-ant with respect to a transformation via appropriate poolingin C cells then recognition performance inherits the invari-ance built into the encoding. The actual templates them-selves do not enter the argument: the set of similarities ofthe input image to the templates need not be high in orderto be invariant. From this point of view, the good perfor-mance achieved with random features on some vision taskscan largely be attributed to the invariance properties of thearchitecture.