The human brain effortlessly categorises objects at different levels of abstraction. An important question for our understanding of visual processing is to what extent this categorisation capacity is underpinned or driven by mid-level visual features (e.g., shapes and texture) that co-vary across different categories (e.g., animals or vehicles). Here we addressed this by training two independent Generative Adversarial Networks to produce two categories of novel synthetic stimuli that match the low- and mid-level visual features of either animate or inanimate real objects. Crucially, where these synthetic stimuli have the low- and mid-level visual properties of real objects, they have no associated high-level category labels. We recorded electro-encephalography and behavioural categorisation responses to the synthetic stimuli along with their real counterparts in a naïve participant sample. Results revealed animacy-like neural signatures for the synthetic objects that emerged earlier than neural signatures distinguishing the same stimuli from images of real objects. The temporal dynamics of neural responses suggest that successful animacy categorisation of our novel synthetic images was driven by their mid-level visual features. These findings contribute to a precise picture of how image statistics support successful object recognition.
