Abstract
Image features computed by specific convolutional artificial neural networks (ANNs) can be used to make state-of-the-art predictions of primate ventral stream responses to visual stimuli.
However, in addition to selecting the specific ANN and layer that is used, the modeler makes other choices in preprocessing the stimulus image and generating brain predictions from ANN features. The effect of these choices on brain predictivity is currently underexplored.
Here, we directly evaluated many of these choices by performing a grid search over network architectures, layers, image preprocessing strategies, feature pooling mechanisms, and the use of dimensionality reduction. Our goal was to identify model configurations that produce responses to visual stimuli that are most similar to the human neural representations, as measured by human fMRI and MEG responses. In total, we evaluated more than 140,338 model configurations. We found that specific configurations of CORnet-S best predicted fMRI responses in early visual cortex, and CORnet-R and SqueezeNet models best predicted fMRI responses in inferior temporal cortex. We found specific configurations of VGG-16 and CORnet-S models that best predicted the MEG responses.
We also observed that downsizing input images to ~50-75% of the input tensor size lead to better performing models compared to no downsizing (the default choice in most brain models for vision). Taken together, we present evidence that brain predictivity is sensitive not only to which ANN architecture and layer is used, but choices in image preprocessing and feature postprocessing, and these choices should be further explored.
