Synthesizing images to map neural networks to the human brain

Abstract

Computational models can be used to generate hypotheses about the brain. In the visual system, this approach has revealed similarities between how natural images are represented in convolutional neural networks and in cortical regions. However, natural images generate highly correlated representations across the hierarchy of model layers, meaning that each brain region will not correspond selectively to a single layer. Because each model layer performs additional image transformations, gaining a more selective mapping between individual layers and regions could reveal how specific algorithmic stages of processing are instantiated in the brain. To enable this mapping, we developed a generative framework for image synthesis that minimizes the similarity of image representational similarity matrices across model layers, aiming to orthogonalize the distances between the patterns of unit activations evoked by the same set of images. With the patterns orthogonalized, the resulting similarity matrix for each layer provides a fingerprint of the unique computational role of that layer. To test this approach, we synthesized 16 artificial images from the Inception-V1/GoogLeNet model and scanned participants with fMRI while they viewed these images repeatedly in random order. Image-specific patterns of voxel activity were used to compute image-by-image similarity matrices across the whole brain with searchlights. Most layers could be mapped to circumscribed cortical regions, and these mappings overlapped less than the mappings obtained with natural images. Given the prevalence of existing fMRI datasets with natural images, we used the synthesis method as a benchmark to develop an alternative residual method that can achieve comparable performance for natural image datasets. These approaches could be extended to other neural network architectures and stimulus modalities for targeted mappings of model computations to the brain and behavior.