Search Papers | Poster Sessions | All Posters
Poster B81 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Predicting brain activation does not license conclusions regarding DNN-brain alignment: The case of Brain-Score
Gaurav Malhotra1, Jeffrey Bowers2; 1SUNY, University at Albany, USA, 2University of Bristol, UK
The Brain-Score benchmark ranks how well DNNs model human "core object recognition" based on how well DNNs predict a wide range of neural and behavioural data. Here we focus on the Brain-Score predictions of IT neural activation and show that good predictions are not a good measure of DNN-IT alignment. We carry out a controlled experiment using the data from Majaj et al. (2015) to understand which visual features drive DNN brain predictions. We show that a good proportion of the neural predictivity score from the dataset are based on the backgrounds of images rather than the objects themselves. This reflects a more general problem of making claims regarding DNN-brain alignment based on correlational studies.
Keywords: Brain-Score neural predictivity object recognition DNN