Abstract
Neurons in successive areas along the ventral visual pathway exhibit increasingly complex response properties that facilitate the robust recognition of objects at multiple locations, scales, and orientations in the visual world. Previous biological models of object recognition have focused on leveraging these response properties to build up transformation invariant representations through a series of feedforward projections. In addition to feedforward projections, feedback projections are abundant throughout visual cortex, yet relatively little is known about their function in vision. Here, we present a model of object recognition that shows how feedback projections can produce considerably robust recognition performance in visual noise. The model is capable of transformation invariant recognition of 100 different categories of three-dimensional objects and can generalize recognition to novel exemplars with greater than 90% accuracy. When the objects are embedded in spatially correlated visual noise, our model exhibits substantially greater robustness during recognition (a 50% accuracy difference in some cases) compared to a feedforward backpropogation model. Thus, the top-down flow of activation via feedback projections can help to resolve uncertainty due to noise based on learned visual knowledge. Finally, in contrast with other biological models of object recognition, our model develops all of its critical transformation invariant representations through general-purpose learning mechanisms that operate via the feedforward and feedback projections. Thus, our model demonstrates how a biologically realistic architecture that supports generic cortical learning is successful at solving the difficult problem of invariant object recognition.