How lab grown neurons could power the future of AI

In 2017, Demis Hassabis and his colleagues at Google’s Deepmind highlighted the commonalities between artificial intelligence (AI) and neuroscience in a paper and noted that better understanding of the human brain could play a vital role in building intelligent machines.

At that time, Hon Weng Chong, a trained medical practitioner, had just wound up his smart medical device startup and was looking to start his next venture. Inspired by the paper, he visited his alma mater, the University of Melbourne, and spoke to some neuroscience experts.

He was told that there was a book coming out of Japan that described how neuroscientists there had managed to train neurons to recognise two separate signals in what’s known as blind source separation that enables one to distinguish another person’s voice from ambient noise.

“That’s the computational problem that neurons are solving and all you have to do is scale up this process and you’ll get a biological computer,” said Chong, who later secured $1m in seed funding from Blackbird Ventures, an Australian venture capital firm, to take the idea further.

In 2019, Chong started a laboratory in what is now Cortical Labs to figure out how to programme biological neurons on a computer chip to perform intelligent tasks. The classic arcade game, Pong, came to mind.

“The first task that Deepmind got their AI to do was to play Pong and little did we know that Elon Musk was also doing the same thing with monkeys,” Chong said. “We spent the last two to three years collecting data for the neurons and how we can programme them to control the paddles in the game to return the ball.”

We are talking to multiple cloud providers at the moment and they're very keen to look at this technology because neurons consume very little energy and generate almost no heat. So, while you're saving energy costs, you also save money on air conditioning

Cortical initially started with embryonic mice cells that provided a good mix of different types of neurons that were needed to perform the task, but as its chief scientific officer was allergic to mice, it switched to stem cells instead.

By sheer luck, even though the research team didn’t know the optimal composition of neurons, Chong said their “estimation proved to be quite useful” and the neurons that they grew from human stem cells outperformed those of mice cells.

When compared to artificial neural networks, Cortical’s biological neurons were also more efficient when they were being trained on Pong than deep reinforcement learning algorithms that power the likes of Deepmind’s AlphaGo, the first computer programme to defeat a professional human player in the ancient game of Go.

“If you think about why we don’t have more robots today, it’s because the training algorithms are so inefficient that if there was an obstacle that stood in the way, the robot would need to sit there and sample for five minutes before it can figure out what to do.

“Another application is cyber security – if you have a language model or anything like that, it will only be good at defending against threats it knows. But hackers are going to use new exploits rather than a known methodology, so you’ll need systems that can respond in real-time,” he said.

But as much as Cortical has had success with biological neurons, it had to refactor or rewrite the Pong game in C whenever it ran different types of experiments.

Realising the need to abstract away the underlying neurons and make them easily programmable using higher-level languages, it built a computing stack comprising a multi-electrode array chip that holds the neurons in a nutrient rich solution, analogue-to-digital and digital-to-analogue conversion chips that pick up and convert electrode activity into signals, along with field programmable gate arrays (FPGA) chips to execute commands.

In April 2023, Cortical closed a $10m funding round led by Horizons Venture, enabling it to speed up its commercialisation efforts. Its technology has even caught the eye of Amazon’s chief technology officer Werner Vogels who visited its labs recently.

“We are talking to multiple cloud providers at the moment and they're very keen to look at this technology because neurons consume very little energy and generate almost no heat. So, while you're saving energy costs, you also save money on air conditioning. There's a strong incentive for them to find more energy efficient models because at the end of the day, the current AI systems are bounded by how much we can push the silicon,” Chong said.

Meanwhile, Cortical still has some work cut out to ensure its neurons perform consistently. Even though its stem cells are from the same source, there are some cells that play better than others.

“If they're genetically the same, what that means is that there must be some protein expression that’s different,” Chong said. “It’s an internal goal for us to tease out gene expression differences between the high-performing cells and the poor-performing ones and see if we can reverse engineer those expressions.”