hide
Free keywords:
-
Abstract:
Biological neural networks present a great variety of intricate topological features, including distinct connectivity patterns for different neuron types that have highly non-trivial spatial arrangements. Such inhomogeneous network topologies could lead to different population dynamics, such as the emergence of slower timescales [1] in networks with excitatory clustering. Moreover, the formation of synapse-type specific neural assemblies has been shown to boost the formation of input selectivity in upstream neurons via synaptic plasticity [2], suggesting a link between network topology and local learning. Here, we investigate the impact of inhomogeneous E/I connectivity on the dynamics of a balanced E/I network and its ability to process multiple inputs simultaneously. In particular, we investigate how synapse-type specific connectivity patterns affect the ability of an echo state network [3] to learn the trajectories of several chaotic attractors simultaneously. We use as a reservoir a sparse recurrent E/I rate network consisting of equal numbers of excitatory and inhibitory neurons. We split the network into four clusters, each receiving the trajectory of one chaotic attractor as input (Fig. 1A) and we evaluate the performance of the network by comparing the distributions of points of the input and predicted time series (Fig. 1B). We investigate how cross-connections between clusters impact the network's dynamics and performance in the prediction task. We focus on whether varying excitatory and inhibitory connectivity levels between different clusters can lead to different outcomes. We find that increased inhibitory connectivity between different clusters leads to an increase in the spectral radius of the recurrent network, which pushes the dynamics towards a more asynchronous, near-chaotic state (Fig. 1D). This feature of the dynamics allows us to use the excitatory vs. inhibitory overlap between different clusters as a control mechanism that shifts the network’s dynamics towards a state optimal for processing the chaotic time series. Our findings indicate that the introduction of neuron type-specific connectivity patterns in recurrent networks can exert a powerful influence on neural dynamics and impact the ability of recurrent neural networks to perform computations and learn complex tasks.
