Why is a raven like a writing desk? The interplay of visual and semantic characteristics of animacy in object perception

Recent studies on non-human primates indicate that high-level regions of the ventral visual pathway are organized around a two-dimensional object space. This space captures the main two dimensions on which objects visually differ as derived from activation patterns in a convolutional neural network trained on object classification. Evidence for a similar object space in humans is limited. In four behavioral studies, we explore how a primary dimension of this space, approximating the animate-inanimate distinction, influences human object perception. Our findings show that when objects' appearance matches their true superordinate category (objects on the “correct” side of object space: animate-looking animals, inanimate-looking objects), people effectively use diverse visual aspects for their discrimination, with minimal interference from semantic relatedness. Conversely, when appearance and semantics clash (objects on the “incorrect” side of object space: animate-looking inanimate objects, inanimate-looking animals), visual information is underutilized or misleading, and semantic relatedness hinders discrimination. Furthermore, categorization of clashing objects is led astray under conditions thought to cut short recurrent processing. We suggest that a visual object space dimension that closely aligns with the animate-inanimate distinction is automatically extracted by the human visual system.