Abstract
Visual reasoning is the process of analyzing visual information to solve problems. A recent study (Kim, et al., Interfocus, 2018) demonstrated that state-of-the-art vision algorithms (i.e. deep convolutional neural networks or DCNs) encounter severe limitations when faced with certain visual reasoning tasks, such as determining whether two items are the same or different (SD tasks), but can successfully accomplish tasks based on spatial relationships between items (SR tasks). Here, we investigated the human neural and cognitive processes involved in SD and SR tasks, with the long-term goal of improving vision algorithms by implementing brain-inspired solutions. We hypothesized that distinct brain oscillations may be involved in the two tasks. We recorded EEG signals of 14 participants performing both SD and SR tasks in the same session. In both conditions, two stimuli were simultaneously displayed for 30ms on opposite sides of the screen. One second later, participants were instructed to report whether the two stimuli were identical (in the SD condition), or whether they were aligned closer to the horizontal or vertical meridian (in the SR condition). The results revealed higher activity in the alpha (8–12Hz) and low beta (13–20Hz) bands from ~500ms post-stimulus until the participants’ response. As behavioral performance differed between the two conditions, we performed a more controlled, between-subject analysis on a subset of participants from each condition, matching the two groups for accuracy; this confirmed that occipital and frontal beta oscillations are significantly higher in SD than in SR conditions. We surmise that beta-band modulations may reflect the involvement of memory processes required to perform the SD task. Ongoing work aims to replicate these results in a within-subject design, equating behavioral accuracy in both conditions. Altogether this result suggests that current DCNs models may benefit from computational strategies that, in the human brain, rely on beta-band oscillatory mechanisms.
Acknowledgement: ERC 614244 P-CYCLES