Human brain cells grown in a lab have learned to play DOOM. Australia-based Cortical Labs put on this impressive display with their CL1 biological computer, which is a really cool device that cradles 200,000 living human neurons on a microchip topped with a multi-electrode array. These neurons, derived from stem cells, float in a nutrient solution and link directly to the chip’s electrodes, which send and receive electrical signals.

The game’s visual stream is transformed into electrical squiggle patterns, which are then sent into the neurons. Enemies approaching from the left may receive one set of signals, while walls or open passageways receive another. The neurons receive these inputs, process them, and fire their own distinctive patterns, which the system maps back to actions: this pattern means move forward, while this one says turn – or shoot, for that matter. Then there’s feedback, more electrical stimulation, which serves as positive reinforcement for successful motions and a smack on the wrist for errors. It all creates a lovely closed-loop mechanism that allows the neurons to learn and improve over time.

Sean Cole did the setup with Python and the Cortical API. He hooked the game to the CL1 (or via their Cortical Cloud platform) and let the neurons about a week to learn the basics of the game. Not surprisingly, they learned the fundamentals of gameplay in approximately a week. They can easily outperform random actions, but they still fall short of matching the skill level of an average human. The character wanders around rudimentary stages, notices foes approaching the rails, and fires away at them, however the real actions appear uncertain and jerky in the video they released.

This is a significant improvement over Cortical Labs’ previous experiment, which involved neurons playing a basic game of Pong. That’s a lot easier than moving a paddle back and forth to hit a ball. The setup for that one was quite large, 800,000 neurons, and it required much more training to get running. DOOM, on the other hand, adds a lot more to the mix, such as 3D environments, different foes, a variety of firearms to pick from, and a slew of real-time decisions to make while under duress. The CL1 has also advanced significantly, now boasting 59 electrodes and ultra-fast response times of less than a millisecond, ensuring that everything runs smoothly.

According to some scientists, biological systems have the ability to pick up specific occupations far faster than standard silicon-based AIs in certain contexts. Unlike traditional AIs, which require backpropagation algorithms, neurons adapt naturally through built-in plasticity. They’re still not sure how they manage to do it all without eyes or a physical body. The only way to find out is to use better training methods, and Cortical Labs is pleased to let researchers and developers get in there and try it out for themselves using the CL1.

Cortical Labs is currently marketing the CL1 to labs studying brain processing, medication efficacy, and other topics that do not require the use of animal models. The video they published demonstrates one innovative technique to combine biological tissue with computer gear. Of course, it’s a long way from replacing the computers we all use, but the fact that you can get neurons to play a game like Doom shows that biological computing may be able to manage highly uncertain and dynamic settings in ways that standard silicon cannot.