Inspecting voxel responses to visual face features using GAN layer activations
We build upon the existing framework created by Dado et al. (2020). We attempt to gain a new perspective on the properties of voxel receptive elds. We do this by recreating and looking at all individual layers of the PGGAN model by Karras et al. (2018). By taking these individual layer outputs and predicting the voxel responses we map the receptive elds of the voxels based on the di erent activation layers. Fitting these mappings yields the de ning receptive eld properties of each voxel. Altogether, we use both neural decoding and encoding to reveal how visual face features are encoded in voxel responses.
Faculteit der Sociale Wetenschappen