Spiking attractor model of motor cortex explains modulation of neural and behavioral variability by prior target information

Abstract

Both neural activity and behavior of highly trained animals are strikingly variable across repetition of behavioral trials. The neural variability consistently decreases during behavioral tasks, in both sensory and motor cortices. The behavioral variability, on the other hand, changes depending on the difficulty of the task and animal performance. Here we study a mechanism for such variability in spiking neural network models with cluster topologies that enable multistability and attractor dynamics, features subserving functional roles such as decision-making, (working) memory and learning. Multistable attractors have been studied in spiking neural networks through clusters of strongly interconnected excitatory neurons. However, we show that this network topology results in the loss of excitation/inhibition balance and does not confer robustness against modulation of network activity. Moreover, it leads to widely separated firing rate states of single neurons, inconsistent with experimental observations. To overcome these problems we propose that a combination of excitatory and inhibitory clustering restores local excitation/inhibition balance. This network architecture is inspired by recent anatomical and physiological studies which point to increased local inhibitory connectivity and possible inhibitory clustering through connection strengths. We find that inhibitory clustering supports realistic spiking activity in terms of a biologically realistic firing rate, spiking irregularity, and trial-to-trial spike count variability. Furthermore, with the appropriate stimulation of network clusters, this network topology enabled us to qualitatively and quantitatively reproduce in vivo firing rate, variability dynamics and behavioral reaction times for different task conditions as observed in recordings from the motor cortex of behaving monkeys.

State-dependent regulation of cortical processing speed via gain modulation

“…Biologically plausible models of metastable dynamics have been proposed in terms of recurrent spiking networks where neurons are arranged in clusters, reflecting the empirically observed assemblies of functionally correlated neurons [29][30][31][32]. Clustered network models of metastable dynamics provide a parsimonious explanation of several physiological observations such as stimulus-induced reductions of trial-to-trial variability [24,25,27,53,54], of firing rate multistability [25], and of neural dimensionality [26]. Our results extend the biological plausibility of clustered networks by showing that they capture other ubiquitous features of cortical dynamics: they operate in the inhibition stabilized regime [55][56][57]; they naturally give rise to lognormal distribution of firing rates [58][59][60][61].…”

“…Our theory is based on a biologically plausible model of cortical circuits using clustered spiking network [22]. This class of models capture complex physiological properties of cortical dynamics such as state-dependent changes in neural activity, variability [23][24][25][26][27] and information-processing speed [20]. Our theory predicts that gain modulation controls the intrinsic temporal dynamics of the cortical circuit and thus its information processing speed, such that decreasing the intrinsic single-cell gain leads to faster stimulus coding.…”

Unsupervised learning and clustered connectivity enhance reinforcement learning in spiking neural networks

“…The input layer projects to the representation layer, consisting of 5000 integrate-and-fire neurons of which 4000 are excitatory and 1000 are inhibitory. This is implemented as a balanced random clustered network, as described in (Rost et al, 2018; Rostami et al, 2020): the connection probability is uniform, but synaptic weights within a cluster (containing both excitatory and inhibitory neurons) are scaled up while the weights between clusters are scaled down, such that the total synaptic strength remains constant (see Figure 1B for a visualization of the network). Thus, the neurons in the representation layer receive input from the input layer and from recurrent connections: where G is the non-linear transfer function of the integrate-and-fire neuron and ϕ is a static background input (bias).…”

“…Our model consists of an input layer, a representation layer and and output layer. The representation layer is based on a balanced random network of spiking integrate-and-fire neurons including clusters as described in Rost et al (2018); Rostami et al (2020). When combined with unsupervised learning (Tetzlaff et al, 2013) of the input projections, the clusters become specialized, in a self-organized fashion, for features of the input space, thus allowing stable representations of the input to emerge that support a linear separation.…”

The generation of cortical novelty responses through inhibitory plasticity

“…It differs from classic attractor models where inhibition is considered unspecific (like a 'blanket') (Amit and Brunel, 1997). Computational work is starting to uncover the functional role of specific inhibition in static networks (Rost et al, 2018;Najafi et al, 2020;Rostami et al, 2020) as well as the plasticity mechanisms that allow for specific connectivity to emerge (Mackwood et al, 2020). These studies have argued that inhibitory assemblies can improve the robustness of attractor dynamics (Rost et al, 2018) and keep a local balance of excitation and inhibition (Rostami et al, 2020).…”