Gillian Rosenberg¹, Caitlyn McGlashan², Benjamin Wolfe³, Michelle Greene¹; ¹Barnard College, ²Bates College, ³University of Toronto Mississauga

Safe driving requires constant assessment of potential hazards. Observers can correctly report the presence of a hazardous road condition in ~400-600 ms (Wolfe et al., 2020), but reaction times combine perceptual processing, decision time, and motor response. While object and scene categorization can be assessed from M/EEG in 150- 250 ms, these studies use well-defined basic-level categories. However, hazards are not a fixed category. For example, a moose is only hazardous when it is located too close to a road. Thus, hazards can be understood as an ad hoc category where category members depend on context (Barsalou, 1983). Is ad hoc hazard categorization slower than basic-level categorization? We presented participants (N=16) with 8-second dash cam videos that either showed a hazard or no hazard at some point in the visual epoch. We used 64-channel EEG to measure brain activity while observers viewed each video. We then performed cross-validated logistical regression analyses at each time point (from 1000 ms before to 1000 ms after the hazard) to determine how well the EEG data predicted the presence of a hazard using a log likelihood metric. To control for the possibility that videos with hazards have more motion, we calculated motion energy for each video and fit a full model using both motion energy and EEG features, as well as a model with only the motion energy. We subtracted the R2 of the motion energy model from the R2 of the full model to obtain the prediction from EEG alone. We found that significant hazard decoding was possible starting at 142 ms after hazard onset. These findings indicate that hazard categorization has the same lower bound as basic-level categorization, showing that the ad hoc and contextual nature of the hazard category does not strongly impact the time course of category processing.

Acknowledgements: NSF CAREER 2240815