Performance-optimized hierarchical models only partially predict neural responses during perceptual decision making

Models of perceptual decision making have historically been designed to maximally explain behaviour and brain activity independently of their ability to actually perform tasks. More recently, performance-optimized models have been shown to correlate with brain responses to images and thus present a complementary approach to understand perceptual processes. In the present study, we compare how these approaches comparatively account for the spatio-temporal organization of neural responses elicited by ambiguous visual stimuli. Forty-six healthy human subjects performed perceptual decisions on briefly flashed stimuli constructed from ambiguous characters. The stimuli were designed to have 7 orthogonal properties, ranging from low-sensory levels (e.g. spatial location of the stimulus) to conceptual (whether stimulus is a letter or a digit) and task levels (i.e. required hand movement). Magneto-encephalography source and decoding analyses revealed that these 7 levels of representations are sequentially encoded by the cortical hierarchy, and actively maintained until the subject responds. This hierarchy appeared poorly correlated to normative, drift-diffusion, and 5-layer convolutional neural networks (CNN) optimized to accurately categorize alpha-numeric characters, but partially matched the sequence of activations of 3/6 state-of-the-art CNNs trained for natural image labeling (VGG-16, VGG-19, MobileNet). Additionally, we identify several systematic discrepancies between these CNNs and brain activity, revealing the importance of single-trial learning and recurrent processing. Overall, our results strengthen the notion that performance-optimized algorithms can converge towards the computational solution implemented by the human visual system, and open possible avenues to improve artificial perceptual decision making.