In 2022, a team based in Melbourne connected a layer of living neurons to a video game and watched them learn to play it. Actual neurons, roughly 800,000 of them, growing on a dish, with no eyes, no body, and no reward system anyone had given them. At least not on purpose.
The setup is called DishBrain, built by a team at Cortical Labs led by neuroscientist Brett Kagan. The neurons were a mix of mouse cortical cells and human neurons derived from stem cells. They were cultured directly onto a high-density array of electrodes, a grid dense enough to both listen to the electrical activity of individual neurons and deliver precise electrical stimulation to them cell by cell.
The team wired this array into a simulated version of Pong. The position of the ball was converted into a stimulation pattern delivered to one region of the dish. The neurons’ own resulting activity in another region was read out and converted into paddle movement.
This is much stranger than a standard reinforcement learning setup because, what were the neurons were working with when they got something wrong?
A game-playing AI is normally trained by maximizing reward, like points for winning or a number going up. Neurons in a dish don’t have access to anything like a dopamine signal you can just hand them. So the team, working with theoretical neuroscientist Karl Friston, a co-author on the paper, built the feedback around Friston’s free energy principle instead, the idea that biological systems act, fundamentally, to minimize surprise about their environment rather than to chase reward directly.
When the dish hit the ball, it received predictable, structured stimulation.
When it missed, it received unstructured electrical noise.
Instead of earning a prize laying well earned a more orderly, a less chaotic world to be stimulated by.
Within about five minutes of gameplay, the culture’s rallies got longer. The neurons weren’t just twitching randomly or occasionally connecting with the ball by chance, their hit rate improved over the course of play, in a way the researchers could track and reproduce.
Interestingly, the human-derived neurons sustained longer rallies than the mouse-derived ones, though the team has been careful not to lean too hard on that difference given how early this work still is.
Kagan himself ALSO stressed that DishBrain shows no evidence of consciousness, and the original paper’s use of the word “sentience” in its title, meaning simply the capacity to sense and respond to an environment, drew real pushback from other neuroscientists who felt it invited exactly the kind of overclaiming Kagan says he wanted to avoid.
A system can behave adaptively, in a measurable, repeatable way, without that behavior implying anything about what it’s like, if anything, to be that system.
Deep reinforcement learning agents typically need enormous numbers of training episodes to master something as simple as Pong. A follow-up study directly comparing biological neuron cultures with deep RL agents on the same task found the neurons could reach comparable performance using far less experience, a property researchers call sample efficiency.
Whatever the neurons are doing to organize themselves around this task, they’re doing it with less data and drastically less power than a GPU cluster needs.
Real translational pull: a living, adaptable neural system on a dish is a plausible platform for testing how drugs, or conditions like epilepsy, alter learned behavior directly, without ever needing an animal model in the loop.
What this suggests, I think, is the capacity to act so as to make your surroundings more predictable isn’t something that waits around for a full brain, a body, or a nervous system shaped by millions of years of evolution to show up. Whatever learning fundamentally is, it may be running on much simpler hardware than we’ve been assuming.
Further reading
Kagan, B.J., Kitchen, A.C., Tran, N.T., et al. (2022). In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron. https://doi.org/10.1016/j.neuron.2022.09.001
Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11, 127–138.
Biological Neurons Compete with Deep Reinforcement Learning in Sample Efficiency in a Simulated Gameworld (2024). arXiv:2405.16946.