Fig. 1: Hierarchical inference in task-independent models.
From: Top-down perceptual inference shaping the activity of early visual cortex
a Example natural image. b Inference in a goal-directed model, which aims to identify categories of images. c Illustration of feed-forward processing in the ventral stream of the visual cortex. d Inference in a hierarchical task-independent model, which permits reverse engineering the contribution of a hierarchical set of features to the observed stimulus. e Illustration of top-down (red arrows) influences supplementing feed-forward processing (blue arrows) along the hierarchy of the ventral stream10. f Computational role of top-down connections. Top: Response intensities of a pair of Z1 neurons (axes) are determined by the features they represent (insets on axes). Interpretation of the image (posterior distribution, empty contours) is the combination of the prior (filled red ellipses) and the evidence carried by the image (likelihood, gray ellipses). Mean response of the neuron (measured across trials, or across a long span of time) is the posterior mean (dot), its correlation is contributing to the noise correlation. Middle: Change of contextual priors upon changes in stimulus statistics. Bottom: Contribution of Z2 neurons to contextual priors in Z1. Saturated colors indicate stronger activation. g Neuronal circuitry of TDVAE that learns a hierarchy of features Z1 (orange) and Z2 (green) corresponding to V1 and V2 in the ventral stream. Features are represented in a layer of neurons (planes with orange and green disks that correspond to individual neurons). Layers of neurons (gray planes) transform feed-forward (Z1-ff, Z2-ff, blue arrows) and top-down information (Z1-TD, red arrow) to ensure precise inference. Integration of feed-forward and top-down information is achieved at Z1-INT. Image credit: (a, f) iStock Photos/Getty Images; (b, d) Users Appaloosa, Patrick Giraud, and VikiUNITED of Wikimedia Commons, user Nikhil Patle of pexels; (c, e) Imaging data provided by the Duke Center for In Vivo Microscopy NIH/NIBIB (P41 EB015897)116,117.
