Australian biotech company Cortical Labs recently posted a video in which 200,000 live neurons grown on a silicon chip played the first-person Doom in 1993. The neuron-controlled protagonist wandered corridors, encountered enemies and fired weapons—incredibly, and died often. But the neurons were playing anyway.
A demo can mark a real inflection point. The neurons appeared to exhibit what Cortical Labs’ chief scientific officer, Brett Kagan, calls “real-time, goal-directed, adaptive learning.” The numbers are expanding beyond the game, in part because AI’s desire for power is growing rapidly. While neurons may not replace microchips, they can perform calculations more efficiently, and analyzing them could provide new ways to use computers—and perhaps even test neuromedicine.
To be clear, Cortical Labs’ neural cells are not extracted from the brain. Kagan explains: “Basically, you can take blood or a small piece of skin, isolate certain types of cells, turn them into stem cells and then, from those cells, produce nerve cells indefinitely.” Each of its computing units can store up to 800,000 neurons in a self-contained system that can keep them alive for up to six months. The structure depends on electricity – “a shared language between biology and silicon,” as he puts it. When brain cells are active, they produce small electrical pulses, and the system can send small impulses back to them.
Supporting science journalism
If you enjoyed this article, consider supporting our award-winning journalism to register. By purchasing a subscription, you help ensure a future of compelling stories about the discoveries and ideas that shape our world today.
But the wiring is the easy part. The hard part is getting the cells into the dish to do anything meaningful. “The temptation is to anthropomorphize and say, oh, they love it [playing Doom],” Kagan says: “But this is not an animal or a person or anything as complex as an insect. It’s a system. It’s like saying, ‘Does the computer like or dislike the pay function [reinforcement-learning] example?’”
The solution to nerve stimulation used the principle of free energy, developed by neuroscientist Karl Friston of University College London. The principle states that nervous systems are driven to predict their environment. Kagan says: “If I can drink an empty can of liquor and predict the consequences of my actions, that’s the kind of world I can live in. But if I reach for it and sometimes it turns into chicken and sometimes it turns into fire, it would be impossible to live in that world.”
To train the neurons, the team built a simple feedback loop. The wrong moves produced unpredictable, unpredictable signals—white noise. The right movements have produced systematic, predictable ones. Kagan says: “Any indication that the cells could not predict is something that the cells would have to learn to avoid, because that would be the only way to create predictability in this environment.” In fact, chaos was the punishment, and order was the reward.
In October 2022 Cortical Labs published a proof-of-concept study in a journal Neuron. Kagan and his colleagues showed that in just a few minutes, neurons in microchips can learn to play Pong, the classic video game in which the player repeatedly builds a ball—think two-dimensional ping-pong. But Pong involves moving squares and moving lines. Doom has corridors, enemies, three-dimensional navigation and a lot of things trying to kill you.
To do this, Cortical Labs organized a hackathon with Stanford University. Independent researcher Sean Cole paired neurons with a common learning method. The hybrid system outperformed the automated algorithm—suggesting that biological cells are involved in the learning process.
Cortical Labs builds its ambitions around two tracks. The first is clinical: “93 to 99 percent of clinical trials, depending on how you cut it, in the neuropsychiatric area fail,” Kagan says. Most of those drugs are tested intravenously, but he points out that brain cells are not designed to live in a place where there is no information. He says: “We’ve actually printed and shown that when you have cells in a sports environment or a world environment, they’re very different in how they respond to drugs, how they manifest in disease.
The second track is computational. Neurons make up “the most powerful information processing system we know,” Kagan says. “Its complexity is beyond anything we’ve built in silicon.” He says that silicon transistors have the first form of complexity – a binary state, 0’s and 1’s. “Biological neurons are at least third-order complex, perhaps much higher. They can hold at least three energy levels at any given time.”
That complexity, the researchers argue, could translate into greater energy savings. Feng Guo, an assistant professor at Indiana University Bloomington, sees Cortical Labs’ biocomputing platform as capable of “high-end computing.” In the 2023 paper Natural ElectronicsGuo and his colleagues developed “Brainoware,” a system that uses three-dimensional brain organoids for computing. For Guo, the power argument is important. The human brain uses only 20 watts—less than a dim light bulb. “If you want to create the same computing power for a silicon-based AI system, it would be a million times more,” he says.
However, Kagan is careful not to overlook the future. He says: “The pocket machine will outlast me in the long division every day. “But your situation is the best. [reinforcement-learning] An AI algorithm is not as good as going into someone’s house and figuring out how to make a cup of tea. ” Biological computing is “a new tool in the intelligence toolbox,” he says.
Don’t expect your computer to be a brain in a vat anytime soon. Kagan makes sense about the research yet to be done but says “you go from science fiction to science once you solve that problem.” A few years ago the biological computer had only one published game of Pong to its name. It now has a commercial platform, an application program interface that developers can connect to a video of neurons tripping in Doom—bad, but still learning.
#human #neurons #chip #learned #play #Doom