Scientists created mini-brains (and taught them how to solve an engineering problem)

A few small masses of lab-grown brain tissue have demonstrated a remarkable proof of concept: living neuronal circuits can be guided to solve a classical control problem through carefully structured feedback.

In a closed cycle system that provided electrical feedback based on performance, cortical organoids managed to progressively improve the control of a classic engineering reference problem: how to balance a unstable virtual pendulum.

The improvement is far from constituting a hybrid biocomputer functional. But as a proof of concept, it demonstrates that neuronal tissue in a laboratory dish can be dynamically adapted through structured feedback — a result that could help study how neurological diseases alter the brain’s plasticity capacity.

The research was presented in a published in the journal Cell Reports.

“We’re trying to understand the fundamentals of how neurons can be adapted to solve problems,” he says. Ash Robbinsrobotics and artificial intelligence researcher at the University of California (UC), in a statement published on .

“If we can understand what drives this process in a laboratory dish, we open new avenues for studying how neurological diseases can affect the brain’s learning ability“, adds Robbins.

O pendulum cart problem is conceptually simple. Imagine balancing a long object, like a ruler or pen, vertically on your open palm.

Unless it’s perfectly aligned, will start to fall. To keep you standing, you need to constantly adjust your hand position as the object sways and wobbles.

Already version of the cart with a penduluma virtual cart can move left or right to keep a hinged pendulum in balance vertical.

The rules are simple and the point of failure is clear: when the pendulum tips too much. But the small mistakes add up quicklymaking it a classic example of a unstable control problem.

The pendulum cart is a problem frequently used in research in reinforcement learning: It is easy to simulate and quick to perform, but unlike pattern recognition tasks, requires constant adjustments and of great precision, rather than a single correct answer.

For Robbins and his colleagues, the pendulum cart represented a new and clear form ofand test the abilities of cerebral organoids.

The organoids were not grown from human tissue, but rather from mouse stem cellscultivated to form small clusters of cortical tissue capable of neuronal signaling.

These organoids were not complex enough to something approaching thought or consciousness, but they were able to send and receive electrical signalsand its internal connections could change in response to external stimulation.

The experience revolved around a cart with a virtual pendulum. Different patterns of electrical stimulation signaled the direction and degree of inclination of the pendulum. The organoids’ responses were then interpreted as forces to the left or right in order to move the cart and counteract the oscillation.

It is important to clarify that the organoids had no understanding of the task. The researchers were testing whether the tissue’s neuronal connections could be adjusted through feedback — that is, whether pulses of electrical stimulation could produce changes that guide the network toward better control.

Every attempt to balance the pendulumdesignated by episodelasted until it exceeded a predefined angle. Performance was monitored in windows of five consecutive episodes.

The organoids were distributed across one of three conditions: no feedback, random feedback provided to selected neurons, or adaptive feedback based on past performance.

The adaptive condition is the most determining. If performance over five episodes was lower than the average over the last 20 episodes, the system would send a brief pulse of high-frequency stimulation. An algorithm adjusted which neurons received these pulses, based on whether similar stimulation patterns had previously been followed by an improvement in control.

To determine whether the organoids were genuinely improvinginstead of simply be luckythe researchers defined a benchmark based on the performance of a completely randomized controller. If the organoid’s best results during a session exceeded what chance could plausibly produce, that session was considered proficient.

As proficiency rates obtained in each of the conditions were remarkable. You organoids without feedback reached the good performance criterion only 2.3% of the time, and those who received random feedback performed well in 4.4% of the time. With continuous adaptive feedback, however, organoids exceeded the proficiency threshold in 46% of cycles.

“When we can actively choose training stimuli, we can effectively shape the network to solve the problem“, says Robbins. “What we demonstrate is ashort term prediction: We can take an organoid in a certain state and transfer it to another state we are targeting, and do it consistently.”

However, “short term” is the right expression. If kept inactive for some time, just 45 minutes, the organoids “forgot” their trainingreturning to baseline performance. Future work could investigate how to improve the organoid’s memory, eventually by increasing its complexity.

“We want to make clear that our goal is to advance brain research and the treatment of neurological diseases, and do not replace robotic controllers and other types of computers by cultured animal brain tissuess in the laboratory”, he says David HausslerUC bioinformatician.

“This could be considered fascinating, but it would raise serious ethical issuesespecially if they were used human brain organoids“, he concludes.