Geometry and dynamics of representations in a precisely balanced memory network related to olfactory cortex

Abstract

Biological memory networks are thought to store information in the synaptic connectivity between assemblies of neurons. Recent models suggest that these assemblies contain both excitatory and inhibitory neurons (E/I assemblies), resulting in co-tuning and precise balance of excitation and inhibition. To understand the computational consequences of E/I assemblies under biologically realistic constraints we created a spiking network model based on experimental data from telencephalic area Dp of adult zebrafish, a precisely balanced recurrent network homologous to piriform cortex. We found that E/I assemblies stabilized firing rate distributions compared to networks with excitatory assemblies and global inhibition. Unlike classical memory models, networks with E/I assemblies did not show discrete attractor dynamics. Rather, responses to learned inputs were locally constrained onto manifolds that “focused” activity into neuronal subspaces. The covariance structure of these manifolds supported pattern classification when information was retrieved from selected neuronal subsets. Networks with E/I assemblies therefore transformed the geometry of neuronal coding space, resulting in continuous representations that reflected both relatedness of inputs and an individual’s experience. Such continuous internal representations enable fast pattern classification, can support continual learning, and may provide a basis for higher-order learning and cognitive computations.

Computational functions of precisely balanced neuronal assemblies in an olfactory memory network

“…S4A ). As observed previously in networks with a single I neuron population 18 , the mean firing rate, the lifetime sparseness, and other response properties were similar between rand and struct networks ( Fig. 4E and Fig.…”

“…These results indicate that precise synaptic balance is important not only to stabilize global activity but also to prevent runaway correlations in recurrent networks with structured connectivity. Moreover, experimental evidence for EI assemblies supports the hypothesis that pDp generates joint representations of odor space and an individual’s experience by confining dynamics to continuous manifolds in activity space 18 .…”

“…Connections between neurons were drawn from a Bernoulli distribution with probability p YX . Synaptic strengths w YX were then fitted to reproduce observations in ex-vivo pDp, as described in 18 . Simulations were performed using Matlab.…”

“…In ISNs, assemblies do not necessarily establish discrete attractor states but ISNs with structured connectivity may exhibit diverse dynamics including chaotic firing patterns and transient responses 27–30 . Nonetheless, ISNs with EI assemblies can represent learned inputs and support pattern classification by confining activity to manifolds in activity space 18 . Generally, recurrent networks with precise synaptic balance can be trained to efficiently perform different computations, conveying a high amount of information per action potential 15,26,31 .…”

“…In autoassociative networks, E/I co-tuning may be established by assemblies including both E and I neurons. In such “EI assemblies”, feedback inhibition tracks the activity of E neurons and, thus, curbs runaway excitation without non-specific network-wide suppression of activity 12,15,18,26 . In ISNs, assemblies do not necessarily establish discrete attractor states but ISNs with structured connectivity may exhibit diverse dynamics including chaotic firing patterns and transient responses 27–30 .…”