A functional assembly framework based on implementable neurobionic material

Abstract

Neurobionic material is an emerging field in material and translational science. For material design, much focus has already been transferred from von Neumann architecture to the neuromorphic framework. As it is impractical to reconstruct the real neural tissue solely from materials, it is necessary to develop a feasible neurobionics framework to realize advanced brain function. In this study, we proposed a mathematical neurobionic material model, and attempted to explore advanced function only by simple and feasible structures. Here an equivalent simplified framework was used to describe the dynamics expressed in an equation set, while in vivo study was performed to verify simulation results. In neural tissue, the output of neurobionic material was characterized by spike frequency, and the stability is based on the excitatory/inhibitory proportion. Spike frequency in mathematical neurobionic material model can spontaneously meet the solution of a nonlinear equation set. Assembly can also evolve into a certain distribution under different stimulations, closely related to decision making. Short-term memory can be formed by coupling neurobionic material assemblies. In vivo experiments further confirmed predictions in our mathematical neurobionic material model. The property of this neural biomimetic material model demonstrates its intrinsic neuromorphic computational ability, which should offer promises for implementable neurobionic device design.

Keywords: decision making; implementable device; neurobionic material; short-term memory.

FIGURE 1

Mathematical model transformation from real to neurobionic assembly. A, Illustration of a connection matrix from a neural assembly. B, Two equivalence and discretization methods for the connection pattern. C, Illustration of the equivalence and discretization for a connection matrix. D, Architecture of an implementable neurobionic device

FIGURE 2

Spontaneous spike probability distribution of a closed neurobionic assembly. A, Random connection matrix with similar excitatory/inhibitory proportions. Neuron number = 1000, excitatory/inhibitory = 1:1. B, Spike‐time series of each neuron in the assembly. Each dot represents a firing. C, Average accumulation‐spike probability curve of the neurobionic material assembly. D, Spike‐time series and related average accumulation‐spike probability curve of the neurobionic material assembly, with or without stimulation inputs. Neuron number = 1000, excitatory/inhibitory ratio = 0.53:0.47. Arrows: neurons of persistent activation; stimulation 1 = stimulation 2 = 50 neurons

FIGURE 3

Spike probabilities cannot be predicted by linear operations. A, Ranking by ascending or descending synapse weight summation cannot sort the spike probabilities. Neuron number = 500, excitatory/inhibitory = 0.53:0.47. B, Weak correlation between the actual spike probability and the expected probability calculated by the synapse weight summation. C, Weak correlation between the spike probability shift and the expected shift calculated by the input summation. D, Swapping the column of the connection matrix cannot maintain the previous spike probability. E, Scatter diagram between exact numerical solution and related spike probability. Neuron number = 400, excitatory/inhibitory ratio = 0.5:0.5, R = 0.98

FIGURE 4

Polarization of the spike probability resulting from an increase in stimulations. A, Spike‐time series of each neuron in the neurobionic material, affected by increases in the stimulations. Neuron number = 500, excitatory/inhibitory ratio = 0.53:0.47. B, Information entropy of the observed neuron spike probabilities from the increase in stimulations. C, Polarization of the spike probability owing to the increase in stimulations

FIGURE 5

Dependence of the activity of neurobionic assembly on the neuron amount and excitatory/inhibitory proportion. A, Spike‐time series of each neuron in the neurobionic material for different neuron amounts and excitatory/inhibitory proportions. B and C, Illustration of the activity regularity for different parameters. D, Relationship between the excitatory/inhibitory ratio and assembly activity. Neuron number = 500

FIGURE 6

Activation of the neurobionic assembly is similar to the real cortex in vivo. A, Immunofluorescent staining of parvalbumin and 4′,6‐diamidino‐2‐phenylindole (DAPI) in different rat cortex (200×). Red: parvalbumin, blue: DAPI. B, Western blots and histograms showing relative levels of GABAR1 and GABAR2 in Tissues A, B, and C. C, Representative in vivo EEG recording for Tissues A, B, and C cortex. D, Theoretical prediction of the activation in neurobionic tissue. Tissue A: normal cortex; Tissue B: FeCl₃‐induced injured cortex with DFO intervention; Tissue C: FeCl₃‐induced injured cortex; n = 6 per group, data are mean ± SD, *P < .05 versus Tissue A, **P < .01 versus Tissue A

FIGURE 7

Phase coupling among artificial neurobionic assemblies. A, Different coupling phases and thresholds leading to different outputs from the same two neurons. B, Connection framework of neurobionic assemblies embodying memory. There are two input units and four coupling units to be connected for phase coupling and recording. C, Spike‐time series of the outcome coupled by four coupling units when three kinds of inputs disappear. The threshold of output unit was adjusted by an increasing threshold modulator to highlight firing distinction. D, Heat maps of spike probabilities of outcome unit, which can show three different firing modes representing three kinds of removed inputs. Distinctions remain under an increasing modulation threshold