Search Papers | Poster Sessions | All Posters
Poster A2 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Recurrent Convolutional Neural Network Predicting Early Visual Recognition on Dynamic Handwriting Images
Sungjae Cho1,2,3 (), Eilif Muller1,2,3; 1Université de Montréal, 2CHU Sainte-Justine Research Centre, 3Mila - Quebec Artificial Intelligence Institute
The ability to accumulate evidence and make timely perceptual conclusions in a rapidly changing environment is important in many ecological contexts. Here we propose a recurrent convolutional neural network (RCNN) that can predict human decision times of early multi-class recognition on dynamic handwriting images. We adapted the original RCNN to perform multiple binary decisions and to have two accept and reject thresholds for each class, to model human uncertainty thresholds. With these modifications, our model achieves high classification accuracy while better predicting human decision times than the original RCNN and model lacking recurrence. Moreover, the uncertainty of our model aligns well with human perceptual ambiguities early in the stimulus sequences. Our modeling results thus support the notion that recurrence is an important component in perceptual decision-making models for dynamic visual stimuli.
Keywords: convolutional neural network recurrent neural network visual recognition decision making