Dishbrain may sound like a schoolyard insult, but it’s the name researchers chose for a new biocybernetic brain—one that is decent at the classic arcade game Pong.
The team cultured human and mouse cortical neurons on specialized computer chips called multielectrode arrays and connected those chips to a computer running a simplified version of Pong. The “player” could move one paddle up and down on the left edge of the screen in hopes of blocking a ball bouncing geometrically around the space. Success or failure at the task changed the feedback the chip sent to the neuron clusters, which slowly got better at the game, keeping the ball bouncing longer and longer (Neuron 2022, DOI: 10.1016/j.neuron.2022.09.001).
The team, led by Brett J. Kagan of the biological computing start-up Cortical Labs, tried several feedback systems. The brains on chips learned best when a failure to block the ball prompted the system to deliver random stimulus. “When faced with unpredictable sensorium,” the researchers write in the paper, “playing ‘Pong’ successfully acts as a free energy-minimizing solution.” In other words, it seems that neurons dislike uncertainty, and we can use that to teach them video games.
“This is both creepy and cool,” says Sagi Eppel, a University of Toronto researcher who studies machine learning and computer vision and was not involved with the work. “Many methods in artificial intelligence are already brain inspired, so these results can inspire new types of algorithms to train artificial neural nets,” he tells Newscripts.
Kagan and his collaborators predict that from these humble beginnings, practical synthetic biological intelligence systems like these could soon outstrip all-silicon AI in usefulness because living biological systems can evolve and adapt.
For historical reference, roughly 20 years passed between when Pong sequels first shipped with an AI opponent option and when IBM’s Deep Blue supercomputer defeated human international chess grandmaster Garry Kasparov.
If the idea of living, learning supercomputers shakes your certainty in humanity’s continued dominance on Earth, Newscripts can offer one consoling result from the paper: human cell lines consistently outperformed mouse cell lines at Pong. So at least we’ll have that going for us.
Before Dishbrain’s progeny overthrow human governments, they’ll need good robotic limbs to interact here in fleshspace. A team at the Korea Institute of Machinery and Materials has a new option for them, one that mimics the action of an elephant’s trunk.
Sung-Hyuk Song and coworkers placed air-powered suction channels in a soft matrix mounted around a pinch-grip armature controlled by wires. The trunk mimic is able to pick up and manipulate objects ranging in size from large cardboard boxes to individual acupuncture needles. After some practice, the human operators were able to use the device to accomplish real-life tasks, including moving potato chips without breaking them, making flower arrangements, and even lighting birthday candles.
Eppel—not involved in this work either—says that some soft grippers that use suction already exist, but the category is less developed than standard mechanical fingers. One of Eppel’s past projects was developing a computer vision system to recognize vessels and other materials in a synthetic chemistry lab. Combining the suction with a pinch grip is a cool application, he tells Newscripts, and could be useful for grabbing fragile items such as glassware.
Please send comments and suggestions to email@example.com.