Current state-of-the-art AI approaches to object recognition still face significant challenges. Foremost among these is efficiency, e.g. the ability to learn new objects using only a few examples, and robustness, e.g. the ability to perform with high accuracy across a wide range of image viewing conditions. We propose a new approach to object recognition which draws on the operating principles used by the biological brain. Research within experimental neuroscience has demonstrated that in the process of object recognition, the biological brain generates separate information about both the shape and appearance of an object, and that both of these separate pieces of information are used. It is currently hypothesized that the generation of separate information about shape and appearance may allow the brain to perform object recognition robustly. Additional work has shown that attention plays a key role in allowing the brain to achieve object recognition efficiently. Our approach integrates these two findings about how the biological brain performs object recognition to guide the design of a new neural network-based approach for object recognition. We believe that by building these two principles into a neural network for object recognition, we can unlock significant gains in performance that will allow artificial vision systems to function with few limitations in the real world.
