Conference on Cognitive Computational Neuroscience

Which features of an image are represented in each area of the visual cortex? Early visual areas are thought to extract primitive features such as edges but as we go deeper into the ventral visual cortex which supports object recognition there is less consensus on what features are represented in the population of neurons. Here, we extract and evaluate hypothesis-driven interpretable intermediate features ranging between pixels and object categories. We revealed interpretable representations in neuronal populations of the mid- and high-level ventral visual stream that progress from being retinally-based to object-centered to coding object identity at the highest stage. Furthermore, object-centered representations came to dominate the neural response over time suggesting a dynamical process. These interpretable features explained the neuronal response at a comparable or sometimes better level than feedforward deep neural networks with millions of parameters that are currently considered the best models of natural object recognition.