I’m presenting a poster at the 2014 Cosyne meeting. My poster is titled: Stable population coding for working memory in prefrontal cortex. In collaboration with electrophysiologists, this project combines data analysis and computational modeling to explore the dynamics of working memory representations in prefrontal cortex and potential neural circuit architectures underlying working memory. My abstract is below:
Stable population coding for working memory in prefrontal cortex
John D. Murray, Nicholas A. Roy, Ranulfo Romo, Christos Constantinidis, Alberto Bernacchia, Xiao-Jing Wang
Working memory requires conversion of brief stimulus-driven signals into internal representations that can be maintained across mnemonic delays of several seconds. In primate cortex, electrophysiological studies find stimulus-selective persistent activity in single neurons as neural correlates of working memory. However, recent studies have highlighted temporal variations in delay activity, at single-neuron and population levels. It remains unclear how neuronal populations maintain memory of stimuli despite complex and heterogeneous temporal dynamics.
To address this question, we applied population-level analyses to hundreds of single-neuron recordings from primate prefrontal cortex, during two classic working-memory tasks: oculomotor delayed response and vibrotactile delayed discrimination. Both tasks demand working memory of analog stimuli, but they differ in stimulus properties, behavioral response, and neural tuning.
We describe the dynamics of neuronal population in terms of its temporal trajectory in a high-dimensional space. Despite strong temporal dynamics and heterogeneity across neurons, we found a low-dimensional subspace, via Principal Components Analysis, in which stimulus coding is stable across the cue and delay. Furthermore, we found that sensory and mnemonic representations are temporally overlapping, rather than sequential. These results, similar for both tasks, suggest working-memory representations are stable and raise the question of which neural mechanisms support working-memory encoding, maintenance, and retrieval.
To explore possible mechanisms, we analyzed the activity of several neural circuit models proposed to explain physiological observations in working-memory tasks, including “attractor” models and “random” networks. Each model could explain some, but not all, of the experimental observations: attractor models show low-dimensional, stable memory but cannot account for dynamic trajectories, whereas random networks show complex trajectories but have high-dimensional memory representations.
Our work demonstrates the potential of population-level analyses for testing computational models of working memory, and suggests that cortical circuits may operate at an interface between order and disorder that currently lacks an appropriate description.