In 2017, Demis Hassabis and his colleagues at Google’s Deepmind published a paper highlighting the commonalities between artificial intelligence (AI) and neuroscience, suggesting that a better understanding of the human brain will play an important role in building intelligent machines. pointed out that it is possible to
At the time, Hon Weng Chong, a trained medical practitioner, had just finished a smart medical device start-up and was about to launch his next venture. Inspired by the paper, he visited his alma mater, the University of Melbourne, to speak with neuroscience experts.
He explains how Japanese neuroscientists train neurons to recognize two separate signals in a method known as blind source separation, which enables them to distinguish other people’s voices from ambient noise. I was told that a book explaining the
“This is a computational problem that neurons are solving, and if you just scale up this process, you can have a biological computer.”, taking the idea further.
In 2019, Chong opened a lab at what is now Cortical Labs to figure out how to program biological neurons on computer chips to perform intelligent tasks. The classic arcade his game, Pong, came to mind.
“The first task Deepmind gave the AI was to play Pong, but few people knew that Elon Musk was doing the same with monkeys,” Chong said. “We spent the last two to three years collecting data on neurons and how to program them to control the paddles and return the ball in the game.”
We are currently in discussions with multiple cloud providers, and they are interested in this technology because neurons consume very little energy and generate very little heat. That means you save on energy costs while also saving on air conditioning costs.
Hon Weng Chong, Cortical Labs
Cortical originally started with embryonic mouse cells that had the right mix of different types of neurons needed to perform a task, but the chief scientist was allergic to mice, so they switched to stem cells instead. I was.
Quite fortunately, the research team did not know the optimal composition of neurons, but said their “predictions proved to be very useful,” and that neurons grown from human stem cells were similar to those of mouse cells. Better than Neuron.
When compared to artificial neural networks, Cortical’s biological neurons were more efficient when trained on Pong than deep reinforcement learning algorithms such as Deepmind’s AlphaGo. It was the first computer program to beat a professional human player at the ancient game of Go. .
“If you think about why there are not more robots today, it’s because the training algorithms are very inefficient, and if there’s an obstacle in the way, the robot has to sit there and sample for five minutes. I understand what to do.
“Another application is cyber security. If you have something like a language model, it’s only good at defending against known threats. will be needed,” he said.
However, much like Cortical’s success with biological neurons, the Pong game had to be refactored or rewritten in C each time we performed different kinds of experiments.
Recognizing the need to abstract the underlying neurons so that they can be easily programmed using a high-level language, the company developed a multi-electrode array chip that holds neurons in a nutritious solution, analog to digital to digital. We have built a compute stack that includes An analog conversion chip that picks up the electrode activity and converts it into a signal, along with a field programmable gate array (FPGA) chip to execute commands.
In April 2023, Cortical completed a $10 million funding round led by Horizons Venture to accelerate its commercialization efforts. The technology even caught the eye of Amazon’s Chief Technology Officer Werner Vogels, who recently visited the lab.
“We are currently in discussions with multiple cloud providers, and we are looking at this technology because Neuron consumes very little energy and produces very little heat. Ultimately, current AI systems are limited by how far silicon can be pushed, so there is a strong incentive to find more energy-efficient models,” Chong said.
Cortical, on the other hand, still has some work to do to ensure its neurons work consistently. Its stem cells come from the same source, but some cells perform better than others.
“If they are genetically the same, what that means is that there must be different protein expression,” Chong said. “It is our internal goal to uncover differences in gene expression between high-performing and low-performing cells and see if we can reverse engineer their expression.”
