Melbourne researchers have embedded lab-grown human brain cells onto silicon chips, opening up a whole new world of possibilities.
A research team from Monash University, in collaboration with Melbourne start-up Cortical Labs, have unlocked a way for human intelligence to give their DishBrain system continual learning capabilities that current AI lacks.
After growing around 800,000 brain cells in a dish last year, the researchers were able to “teach” the cells to perform goal-directed tasks, such as the tennis-like computer game Pong.
According to project lead Associate Professor Adeel Razi, the idea came about because the researchers wanted to “exploit and control the computational power of a human brain.”
With a background in electrical engineering and a PhD in wireless communication, and now as a computational neuroscientist, Razi said that there are a number of issues with large language models (LLM) such as Open AI’s ChatGPT.
“The problem with current AI technology is that you train [tools] with huge amounts of data and energy, and they can do one specific task very well,” Razi told create. “But then you train them on another task, again consuming large amounts of data and energy, but they will forget what they’ve previously learned. We call this catastrophic forgetting.”
Humans, on the other hand, learn new tasks and skills, even very different ones, without forgetting what we learned in the past. It’s called lifelong, or continual learning, meaning that if we learn to swim we don’t forget how to walk. That’s because our brains are extremely flexible.
“Humans can learn a huge amount with a very small amount of data, and we’re very energy efficient,” Razi said. “However, we don’t entirely know what the brain mechanisms of lifelong learning are.”
How it works
In 2018, Razi and his colleagues at Cortical Labs set about trying to use human brain cells as a biological computer instead of traditional entirely silicon-based platforms.
Culturing human brain cells in the lab with induced pluripotent stem cells (iPSCs), the researchers were able to use stem cell technology to turn cells from an arm’s skin into cortical cells.
“We take a very high-density multielectrode array, which contains 26,000 silicon-based electronic channels using CMOS technology,” Razi said. “Then we plate the living brain cells on this electronic surface or the dish.”
Once the connection between the cell cultures and the dish surface is established, researchers make use of the Cortical Lab’s low latency, real-time feedback loop. Neurons on top of this electronic surface can be stimulated, allowing the researchers to read the activity that they generate.
“The cells respond to electrical signals from the electrodes in the dish,” Razi explained. “These electrodes both stimulate the cells and record changes in neuronal activity. The stimulation signals and the cellular responses are converted into a visual depiction of the Pong game.”
Razi and his colleagues showed that, after the first five minutes of playing Pong, the cells got better at the game over the next 15 minutes.
Silicon computing is not adaptive in this way. Razi said that in future, DishBrain, though still in its infancy, has the potential to outperform only silicon-based computers — especially for use cases that require flexible behaviour.
Bringing it all together
Razi said that bringing together disparate fields, in biology, lab sciences, electronics and software engineering, was one of the greatest engineering challenges.
On a technical level, the multielectrode array is configured to read up to a particular 1024 of its 26,400 electrodes, at a rate of 20,000 samples per second.
Because neurons fire at a rapid rate, the engineers needed to make the system that records and decodes data extremely quickly. Capturing all this information in low-latency, real-time feedback is a feat of electrical and software engineering.
Understanding the biological mechanisms of lifelong learning will help researchers to develop better AI technology with much improved learning capability.
Over the next three to five years, Razi believes that the breakthrough will help researchers gain a “mechanistic understanding” of various brain functions. This will also help them understand the brain’s dysfunction with pathologies such as dementia, and improve drug discovery.
In the longer term, he thinks we’ll see it lead to innovative brain machine interfaces — devices such as robotic limbs that can be controlled with our thoughts — and improve current computing through efficiency and higher throughput resembling the human brain.
Simon Longstaff, Executive Director of The Ethics Centre, said that the research breakthrough raises a cluster of ethical issues for people donating their cells for the technology.
“Would the benefit flow back to the person who’s ultimately provided the material that has gone into the system that’s been deployed?” he asked.
In the future, there may also be concerns about the material being exposed to pain, rather than reward. There’s also what Longstaff calls “the science fiction end” of the technology reaching some form of sentience.
“Do we understand the gradations in these things?” he asks. “And what do we do to guard against taking steps which could cross any of the ethical barriers that would arise?”
Longstaff said that governments are still playing catch up on how to effectively regulate the burgeoning field of AI. He believes that groups such as the CSIRO’s Responsible AI Network are a good start.
“The important thing to get into the mix here is how you bring the ethics to bear without stifling innovation,” he said.
Read how a computer chip developed by RMIT researchers in 2019 can make – and erase – memories like a human brain.