ABSTRACT
We organize and interpret our visual world in terms of objects. Object vision seems effortless, but it is one of the brain’s most daunting information processing tasks. This is because early visual representations are enormously high-dimensional (detailed images distributed across millions of neurons), they contain no explicit information about objects, and they are extremely unstable due to continual changes in position, viewing angle, lighting, occlusion, and other factors. This highdimensional, implicit, unstable input representation, coupled with the virtually innite span of object space, presents a unique coding challenge. Subsequent processing in higher-level visual cortex must transform image information into a very different representation that can support our remarkable capacity for object perception and object memory. Object information must be compressed into a much lower dimensionality (i.e., a reasonable number of neural signals) that can be transmitted to other parts of the brain and stored efciently in memory. High-level object representations must be explicit-that is, there must be a simple, consistent relationship between neural signals and object characteristics-so that they can be rapidly decoded and utilized by other parts of the brain. Highlevel object representations must also be stable across image changes at the retinal level-changes in retinotopic position due to eye movements, changes in size due to distance, and changes in orientation relative to the observer. Finally, a high-level object code must be productive-it must be able to generate a huge number of distinct patterns in order to represent the entire world of natural and articial objects.
