Low-dimensional prefrontal representations of objects during working memory
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
There is an ongoing debate regarding the dimensionality of neural representations. Some accounts emphasize representation within low-dimensional subspaces or manifolds, while others suggest high-dimensional neural codes where neurons respond independently. Here, we investigate the dimensionality of prefrontal cortex (PFC) representations of visual objects held in working memory. We found that object representations are low-dimensional, occupying only 3–6 effective dimensions during both encoding and maintenance in working memory. Control analyses indicate this dimensionality was not limited by the number of objects tested (40) or neurons sampled (∼100). We also compared PFC dimensionality to that of a well-established deep neural network model of its inferotemporal (IT) inputs and found an approximately 7-fold dimensionality reduction in PFC. These results suggest object representations are compressed into a low-dimensional manifold in PFC, which might be related to attractor dynamics for working memory, and might facilitate interaction with other variables and cognitive control.
Author summary
Competing accounts of cortical neural coding propose that neurons either activate in structured, coordinated patterns like schools of fish, or that they activate independently, like hunting sharks. A key metric discriminating these schemes is their effective dimensionality — the number of independent activity patterns needed to characterize information conveyed in a neural population. To test this, we measured the dimensionality of primate prefrontal cortex during working memory for objects, a domain where the inputs are intrinsically complex and high-dimensional. We found that prefrontal dimensionality was surprisingly low, spanning only a handful of dimensions. Using convolutional neural network models of its visual cortex inputs, we found evidence for a substantial prefrontal compression of dimensionality. We hypothesize that this reformatting of high-dimensional sensory information into a compact code can facilitate flexible combination with information from other domains, as proposed by “mixed selectivity” theories of prefrontal coding.
