Goal directed closed-loop training of network bursts in cortical cultures produces improved performance and error correction over days

Abstract

Dissociated neural cultures are living systems with no evolutionary history in their current form, which makes them a testbed for asking what goals neural tissue can learn to pursue once it is removed from its evolved body. It remains unclear which features of network activity such a culture can be trained to control, and whether that training is retained from one day to the next. We built an open-source closed-loop platform, the Holistic Agential Learning platform (HAL), that detects a network burst on a high-density microelectrode array, estimates its spatial origin in real time, and applies a minimal reinforcement rule. Here, a burst that originates off the goal side triggers global random stimulation that persists until the next burst. Over days of training, rat cortical cultures (n=10) increased the fraction of bursts that initiated on the goal side. This gain accrued between sessions, with little change within a session, a consolidation-like pattern that we interpret as a statement about timescale while the mechanism remains untested. The burst origin moved toward the goal side, while the burst peak and the repertoire of propagation directions were preserved. Trained cultures also recovered from incorrect states more reliably over days, while the rate at which they entered those states was unchanged, and this recovery effect was the most robust difference from unstimulated controls (n=6). Cultures that began on either side of chance converged to a similar level of performance. Damaged cultures (n=6) with comparable burst statistics showed neither the across-day gain nor the consolidation-like pattern. Training reweighted the activity patterns each culture already had, and we saw no new ones, which we read as a simple form of niche construction in which the tissue keeps a trace of its history that biases what it does next. These results identify burst initiation and error recovery as trainable variables in a novel living system, and they support training the tissue rather than engineering the machine.