Spiking recurrent neural networks represent task-relevant neural sequences in rule-dependent computation

Abstract

Prefrontal cortical neurons play in important roles in performing rule-dependent tasks and working memory-based decision making. Motivated by experimental data, we develop an excitatory-inhibitory spiking recurrent neural network (SRNN) to perform a rule-dependent two-alternative forced choice (2AFC) task. We imposed several important biological constraints onto the SRNN, and adapted the spike frequency adaptation (SFA) and SuperSpike gradient methods to update the network parameters. These proposed strategies enabled us to train the SRNN efficiently and overcome the vanishing gradient problem during error back propagation through time. The trained SRNN produced rule-specific tuning in single-unit representations, showing rule-dependent population dynamics that strongly resemble experimentally observed data in rodent and monkey. Under varying test conditions, we further manipulated the parameters or configuration in computer simulation setups and investigated the impacts of rule-coding error, delay duration, weight connectivity and sparsity, and excitation/inhibition (E/I) balance on both task performance and neural representations. Overall, our modeling study provides a computational framework to understand neuronal representations at a fine timescale during working memory and cognitive control.Author SummaryWorking memory and decision making are fundamental cognitive functions of the brain, but the circuit mechanisms of these brain functions remain incompletely understood. Neuroscientists have trained animals (rodents or monkeys) to perform various cognitive tasks while simultaneously recording the neural activity from specific neural circuits. To complement the experimental investigations, computational modeling may provide an alternative way to examine the neural representations of neuronal assemblies during task behaviors. Here we develop and train a spiking recurrent neural network (SRNN) consisting of balanced excitatory and inhibitory neurons to perform the rule-dependent working memory tasks Our computer simulations produce qualitatively similar results as the experimental findings. Moreover, the imposed biological constraints on the trained network provide additional channel to investigate cell type-specific population responses, cortical connectivity and robustness. Our work provides a computational platform to investigate neural representations and dynamics of cortical circuits a fine timescale during complex cognitive tasks.

A geometric framework for understanding dynamic information integration in context-dependent computation

“…Classical and new theoretical models in WM have been reviewed and tested in light of recent experimental findings (Lundqvist et al, 2016;Miller et al, 2018). Similar to other modeling efforts (Orhan and Ma, 2019;Xue et al, 2021), the single-unit response in our computer simulations showed strong temporal dynamics rather than persistent activity during the delay epoch. Therefore, our model supports the finding of experimental studies that information is often stored in dynamic population codes during WM.…”

“…Goudar and Buonomano (2018) demonstrated that time-varying sensory and motor patterns can be stored as neural trajectories within the RNN, helping us understand the time-warping codes in the brain. RNNs have also been incorporated with more biological features, such as Dale's principle, which serve as a valuable platform for generating or testing new hypotheses (Song et al, 2016;Xue et al, 2021;Rajakumar et al, 2021).…”

Excitatory-Inhibitory Recurrent Dynamics Produce Robust Visual Grids and Stable Attractors

“…We trained biologically-constrained excitatory-inhibitory (E/I) RNNs (Song et al 2016; Rajakumar et al 2021; Xue et al 2022) to perform a spatial navigation task in a two-dimensional (2D) environmental enclosure. We envisioned that the RNN received various forms of visual and spatial cues in the input (Table 1), and predicted the position.…”

Biophysical control of neural trajectories under disordered brain dynamics

“…It is however unclear how exactly such highly complex synaptic architectures are learned. Other theoretical studies have addressed how neural trajectories embedded in noisy dynamics may emerge from training recurrent neural networks with artificial learning rules 12–14 . However, from the biophysical standpoint, a more plausible candidate for learning neural trajectory oriented connective pathways is the Spike-Timing Dependent Plasticity (STDP) rule 15 .…”