Emergent organization of multiple visuotopic maps without a feature hierarchy

Abstract

The primate visual system is comprised of multiple visual areas. Despite their foundational relevance, there are no normative accounts for why there are multiple areas nor why they have their signature “mirrored map” topography. Here I show that the stereotyped cortical organization of multiple mirrored areas naturally emerges in simulated cortex, in which self-organizing processes are used to map a multi-scale representation of visual space smoothly onto a two-dimensional cortical sheet. Predominant accounts of multiple areas emphasize hierarchical processing, where each area extends and elaborates on the previous areas’ representation. Here, no explicit hierarchical relationships were required to manifest this multi-areal organization, suggesting that feature hierarchies may be the derived rather than the driving force of this organization. This modeling work thus provides a simple computational explanation for the hallmark features of early visual topography, and the presence of multiple areas, as emergent from a single functional goal — to smoothly represent the visual field at multiple spatial scales.One Sentence SummaryThis work presents a formal model of simulated cortex with multiple visual areas, where purely spatial relationships underlie the large-scale motifs of visual cortex.

Brain-optimized neural networks learn non-hierarchical models of representation in human visual cortex

“…We interpret this work as a corrective to a tendency to treat "hierarchy" as the defining feature of primate visual organization. Our interpretation is consistent with emerging evidence for non-hierarchical representation in other sensory systems [Hamilton et al, 2021], motivates the development of networks with multi-branch architectures [Bakhtiari et al, 2021] that model parallel visual streams with distinct functions [Ungerleider, 1982, Schiller and Logothetis, 1990, Pitcher and Ungerleider, 2021, and underscores the need for computational models that treat hierarchy as an emergent property [Konkle, 2021] rather than a requirement for successful vision.…”

Visuospatial coding as ubiquitous scaffolding for human cognition

“…A mechanistic account for why the brain initially recapitulates the retina across its surface is that it minimizes biologically expensive long-range connections (the wiring optimization principle; [32]). Continuing this reference frame throughout visual cortex may be advantageous because it reduces the need for transforming information, allowing distinct aspects of a visual image to be computed in different maps while preserving spatial arrangements [2] (but see [33] for an alternative explanation for repeating retinotopic maps). Having visuospatial codes in visuomotor regions responsible for generating eye movements (e.g.…”

Principles governing the topological organization of object selectivities in ventral temporal cortex

“…In contrast, the SOM model operates in an opposite way by mapping a 2dimensional lattice into the input space (Durbin & Mitchison, 1990;Konkle, 2021), while preserving the most important topological and metric relationships of objects in the high-dimensional object space. The SOM model assumes a regularly, locally connected cortical surface and learns an embedding in the input space, which makes it well-suited to answer the question of how a 2-dimensional cortical surface enables the fundamental ability to represent a multi-dimensional object representation space in our study.…”

“…therefore, the SOM model is a biologically plausible model that can achieve local wiring cost minimization (Durbin & Willshaw, 1987;Kohonen, 1989). Though the SOM models have been successfully applied to predict the appearance of ocular dominance, orientation preference, and direction preference maps in the early visual cortex (Bednar & Miikkulainen, 2003;Durbin & Mitchison, 1990;Konkle, 2021;Linsker, 1988), our study is the first attempt to use the SOM to model the functional organization of higher visual cortex, the VTC, both anatomically and functionally.…”

“…There are two essential features of our computational model. First, unlike previous SOM models that usually take as input the stimulus space (or pixel space) to simulate the lower visual cortex (Bednar & Miikkulainen, 2003;Durbin & Mitchison, 1990;Durbin & Willshaw, 1987;Kohonen, 1989;Konkle, 2021;Linsker, 1988), our SOM model takes a high-dimensional object representation space as the input and outputs a tuned map (i.e., an artificial cortical surface of the VTC), in which nearby units in the map project to nearby points in the object space (Figure 1). The highdimensional object representation space is obtained via a pre-trained deep convolutional neural network (DCNN), the AlexNet, because numerous studies have demonstrated a striking similarity in the response profile and functional hierarchy between the AlexNet and human ventral visual cortex (e.g., Cichy, Khosla, Pantazis, Torralba, & Oliva, 2016;Guclu & van Gerven, 2015;Khaligh-Razavi & Kriegeskorte, 2014;Liu, Zhen, & Liu, 2020;Wen, Shi, Zhang, Lu, Cao, & Liu, 2018;Yamins, Hong, Cadieu, Solomon, Seibert, & DiCarlo, 2014).…”