Fig. 1: Texture and object image manifolds as parameterized by DeePSim and BigGAN.
From: Neuronal tuning aligns dynamically with object and texture manifolds across the visual hierarchy
a, Most visual neurons respond strongly to sets of natural images with little semantic relation. Each curve shows responses to randomly sampled images (top, most activating; bottom, least activating). Shaded areas: s.e.m. b, Architecture and image statistics of the two generative models. (i) DeePSim uses an up-convolutional architecture that generates texture-like, less photorealistic images with a 4,096-dimensional latent space. (ii) BigGAN combines class and noise embeddings into a 256-dimensional latent space and produces object-centric, photorealistic images. Each generator defines a continuous image manifold mapping latent codes to naturalistic images. c, Conceptual framework: neuronal firing rates guide searches across generative manifolds to locate response maxima. Parallel optimizations in DeePSim (texture manifold) and BigGAN (object manifold) reveal complementary aspects of neuronal tuning across the visual cortex.
