Adaptive Goal-Directed Behavior in Embodied Cultured Networks: Living Neuronal Networks and a Simulated Model

Abstract

The advanced and robust computational power of the brain is shown by the complex behaviors it produces. By embodying living cultured neuronal networks with a simulated animal (animat) and situating them within a simulated environment, we study how the basic principles of neuronal network communication can culminate into one of these behaviors: adaptive goal-directed behavior. We engineered a closed-loop hybrid system in which a cultured network controls an animat with a specific sensory-motor mapping and training algorithm. Real-time performance-based feedback allowed both living and model embodied neural networks to learn to move the animat in a desired direction. This approach may help instruct the future design of artificial neural systems and of the algorithms to interface sensory and motor prostheses with the brain.

Contribution of GABAergic interneurons to the development of spontaneous activity patterns in cultured neocortical networks

“…This is of interest because, first, the immature network patterns (cENOs) recorded in slices of the cerebral cortex occur in absence of GABAergic signaling (Garaschuk et al, 2000; Allene et al, 2008). Secondly, although cultured networks have been extensively used as models for cortical network studies (Feinerman et al, 2007; Le et al, 2007; Rolston et al, 2007; Bakkum et al, 2008; Baruchi et al, 2008; Chao et al, 2008; Pasquale et al, 2008; Raichman and Ben-Jacob, 2008; Shahaf et al, 2008), generating cultured cortical networks with a predictable activity pattern development has been difficult (Wagenaar et al, 2006a). Finally, although early synchronized network activity is described as extremely robust behavior, due to homeostatic regulation of intrinsic features (Marder and Goaillard, 2006; Blankenship and Feller, 2010), the limits of homeostasis might be reached along maturation if the local circuitry is abnormally constructed.…”

RT-Sort: an action potential propagation-based algorithm for real time spike detection and sorting with millisecond latencies

“…This important feature of RT-Sort opens up interesting opportunities for closed loop experiments, where a neural system receives external inputs based on preceding activity recorded from the same system. These types of experiments have been performed in various forms but always utilize multi-unit activity as read-out of the neural system (DeMarse et al 2001, Wagenaar et al 2005, Bakkum et al 2007, Kagan et al 2022). With RT-Sort, closed loop experiments can now be conducted with real time single unit read out accuracy, which enables studying neural circuit dynamics in response to external inputs with a greatly enhanced granularity.…”

“…This issue will impact anyone interested in performing closed loop experiments in which a neural system receives external inputs based on preceding activity recorded from the same neurons. Such closed loop experiments have been performed in multiple forms, ranging from controlling the bursting activity of in vitro cultures (Wagenaar et al 2005) to embedding a neural culture in a virtual environment (DeMarse et al 2001, Kagan et al 2022) or letting a neural culture control a robot in the real world while providing information about the robot’s environment (Bakkum et al 2007). Similar issues are faced when reading out signals of brain machine interfaces in real time (Patil et al 2008).…”

MEART: The semi-living artist

“…For the simulated animat (Bakkum et al, 2007 ), the training algorithm was modified in two ways. A pool of candidate PTSs was formed by pairing the probe electrode with other electrodes ( N E = 58) and inter-pulse intervals {−80, −40, −10, 10, 40, 80 ms} ( N PTS = 58 × 6).…”