Subspace alignment as a mechanism for binding

Abstract

To choose between options, we must solve two important binding problems. First, the features that determine each options’ values must be appropriately combined and kept separate from the corresponding features of other options. Second, options must be associated with the specific actions needed to select them. We hypothesized that the brain solves these problems through use of aligned (for bound dimensions) and orthogonal (for separated dimensions) population subspaces. We examined responses of single neurons in six putative value-coding regions in rhesus macaques performing a risky choice task. In all areas, single neurons encode the features that define the value of each option (stakes and probability) but only very weakly encode value per se. However, the coding dimensions associated with these features are aligned on a single subspace, from which a strong emergent value signal can be read out. Moreover, all six regions use nearly orthogonal subspaces for the left and right options, thereby linking options to their position in space, implementing functional partitioning, and reducing the possibility of misbinding. These results provide a new solution to the neuroeconomic binding problems and suggest that other forms of binding may work through similar principles.

Abstract representations emerge naturally in neural networks trained to perform multiple tasks

“…Further, experimental work in the hippocampus and prefrontal cortex has shown that representations of the sensory and cognitive features related to a complex cognitive task, also support generalization 8 . We refer to representations of task-relevant sensory and cognitive variables that support generalization-like in these examples and others [12][13][14][15][16] -as abstract representations.…”

Orthogonal but linked neural codes for value

“…We suspect that two main factors differentiate our study from past ones. First, by focusing on qualitatively different types of offers/choices, we bypass the potential confusion caused by mixed selectivity in prefrontal neurons, which can inflate estimates of pure value coding (Fine et al, 2021). Mixed selectivity has long been known, but its implications, and its powerful ability to alter conventional interpretations of physiological analyses, has only recently come to be appreciated (Rigotti et al, 2013;Fusi et al, 2016).…”

The gradient model of brain organization in decisions involving ‘empathy for pain’

“…These representations may be the furthest removed from the encoding of specific stimuli, as they may be directly translated into motor responses (Cisek, 2012), and are therefore candidates for being hosted at the high end of the cortical gradient (Viviani, Dommes, Bosch, & Labek, 2020). Furthermore, they emerge in a distributed network that integrates signals from different parts of the brain into a low-dimensional criterion (Cisek, 2012; Yoo & Hayden, 2018; Fine, Yoo, Ebitz, & Hayden, 2021). In the current study, participants were presented with two visual scenes and were asked to rate them with the question “which situation is more painful?” We assumed that evidence for decisions was large when one scene represented a painful situation and the other represented a non-painful or neutral situation (‘mixed trials’).…”

“…These representations may be the furthest removed from the encoding of specific stimuli, as they may be directly translated into motor responses (Cisek, 2012), and are therefore candidates for being hosted at the high end of the cortical gradient (Viviani, Dommes, Bosch, & Labek, 2020). Furthermore, they emerge in a distributed network that integrates signals from different parts of the brain into a lowdimensional criterion (Cisek, 2012;Yoo & Hayden, 2018;Fine, Yoo, Ebitz, & Hayden, 2021).…”