Biological neurons are emerging as attractive candidates for artificial intelligence and machine learning applications given their natural energy efficiency and self-repair capacity. However, they differ from their idealized artificial counterparts. Biological neurons have highly variable and noisy dynamics and display intrinsic spontaneous activity instead of purely input-driven dynamics. Moreover, biological neuronal networks have physically constrained and highly plastic connections, leading to a complex and ever evolving connectivity structure. Here, we investigate (numerically and with preliminary experimental data) the stability of the input responses of neuronal cultures using a reservoir computing framework. Utilizing a numerical model for the growth and activity of neuronal cultures, previously used to model experimental data, we investigate the effect of large-scale network topology, specifically homogeneous vs modular architectures, on fading memory, reservoir performance under increasingly noisy dynamics, and robustness to network rewiring. We find that modular networks exhibit longer fading memory time, sustain higher performance under noisy conditions, and are more robust to connectivity rewiring than homogeneous networks. Finally, we observe no relationship between some characteristics of the network adjacency matrix (specifically its spectral properties) and reservoir computing performance.
Houben et al. (Sun,) studied this question.