If you thought the recent surge in AI development was an exotic state of computing with big and topical acronyms like Large Language Model (LLM), Generative Adversarial Network (GAN), think again. . There is nothing more exotic than a human brain cell that can float in a petri dish and work with electronics and his AI agents. But that’s exactly what an Australian team in partnership with Monash University and the Cortical Institute accomplished last year with the introduction of the DishBrain concept.
Apparently, the team’s research concept and promise resonated with the military, and even added dollar value to its appeal as a research facility. US$407,000 military investment in research, courtesy of Australia’s ONI (National Intelligence Agency).
When we introduced DishBrain last year, researchers described a semi-biological processor constructed from a mixture of human and mouse brain cells grown on control electrodes. These electrodes acted as a sort of her BCI (brain computer interface), allowing scientists to send control signals and read out biological “protobrain” activity. Interestingly (although the adjectives may vary), DishBrain was able to demonstrate something akin to perception (a basic survival skill, or minimizing negative feedback) within five minutes of powering on. demonstrated a near-biological duty). “Learn” how to play Pong.
The way they “trained” the brain was just ingenious. The movement of the ball was transmitted to the cells via electrodes, and electrical stimulation of the cells in the dish brain affected different regions mapped to the position of the ball. DishBrain was then given control of the paddle and the ability to move it left and right.
DishBrain suggests this because biological systems tend to consume as little energy as possible (or at least to avoid expensive or negative experiences that reduce energy reserves, i.e. viability). We have set up a simple reward system that we utilize. When the paddle catches the ball, BrainDish’s electrodes give her a “positive”, predictable one-second electrical stimulus to the ball. However, every time the paddle missed a ball, Dishbrain received an unpredictable blast of electricity lasting up to four seconds. In this case, sensation is achieved through viability. Dealing with an unpredictable 4-second burst requires more energy than dealing with a predictable 1-second burst. So DishBain increases survivability by actively maximizing the number of times the paddle hits the ball.
If this conjures up images of someone being forced with an electric baton to appease a particular behavior, you’re not entirely wrong.
Adeel Raj, an associate professor at the Turner Institute for Brain and Mental Health, said their research “brings together the fields of artificial intelligence and synthetic biology to create a programmable biological computing platform.” This promise of having the ultimate parallel processor (biological brain, or wetware) as his computing device is what got him started. Razi said there is even hope that such bioprocessors may eventually surpass the limits of traditional silicon chips.
Interestingly, the DishBrain project has several goals that seemingly align with the current AI arms race. Just as AI companies such as Microsoft and OpenAI are chasing the dream of recursive training (which essentially means AI agents can be trained). itself), DishBrain’s ability to self-learn is the very trophy pursued by silicon and quantum computing-based companies (in Microsoft’s case). Whether that learning is based on biomechanical and physics-based survival instincts or pure perception is a related question, but the end result is the same: the system can learn by itself. .
This ability in particular can have a big impact on AI. Because a processing system based on wetware, rather than current silicon-based approaches, could circumvent one of the biggest stumbling blocks to AI evolution: the ability to remember training, in other words. It’s from “Catastrophic oblivion” that generally prevents AI from building a learning foundation.
According to Raj, the team “will be using this” [the grant] Building better AI machines that replicate the learning capabilities of these biological neural networks. Even though leading researchers are wary of the existential risks posed by AI, research is underway that sounds more likely in a sci-fi context than on the frontlines of the news. tom’s hardware. 2001 movie under And a sentient killer elevator powered by wetware comes to mind.
It’s somewhat stressful to think about what the technology described in this article could do when used for something more military-oriented, such as a drone system that chooses to kill human handlers.
Luckily, only future versions of us will have to deal with it when that moment comes.
