Art by Malia Kuo.
Our brains are a collection of billions of neurons, firing in synchrony to make up the complex organ that is our brain. But zooming in, what if we consider a small, isolated subset of cells? What might they be capable of?
Neurons are unique cells in the body. Unlike other cells, which can simply maintain their functions isolated in a petri dish, neurons process information, meaning they need a stimulus that prompts them to act. This makes them both fascinating and difficult to study. Computational models have been used to study neural networks, but they are limited by the constraints of technology, which is no substitute for a biological system. To alleviate this concern, what if neural networks could be made from biological neurons in a petri dish? In their recent paper, Brett Kagan, his colleagues at Cortical Labs, and several university collaborators have set out to study the interface of biology and intelligence by exploring how biological neurons respond to electrophysiological input and feedback in vitro.
The team’s research process started in 2019. However, the pandemic threw a wrench in their plans, especially with Australia’s strict lockdown procedures. “Fortunately, we were able to get exemptions to go to work because we were considered critical workers, being in a hospital setting. But it was incredibly different circumstances nonetheless, getting supplies in and all the basic little things that we used to take for granted,” Kagan said. Once the lab could work around the restrictions, the group hit the ground running, resuming their research skillfully and deliberately. Kagan and his colleagues adopted an approach called “agile science,” where they set up a series of small pilot experiments in tandem to see which conditions would be best for their cells. This allowed them to adjust their research environment as they went along and optimize their experiments throughout the process. By growing long-term cortical neurons that formed dense connections with supporting glial cells, Kagan and his colleagues were able to study the behavior and capabilities of these biological neurons in a petri dish.
A silicon chip inside the dish stimulated the neurons to create a simulated Pong game-world, where a paddle is moved up and down the screen to block a ball from hitting the side. “We chose Pong because we wanted something [in] real time, simple to code for with a clear ‘win and/or lose’ condition—in this case, there was a really clear lose condition—and [something] recognizable to people. It’s actually the fiftieth anniversary of Pong [this year],” Kagan said. Inputs from the silicon chip were delivered to a predefined sensory area of eight electrodes. These electrodes stimulated sensory neurons that then communicated with motor neurons also cultured in the dish. The researchers wanted to see if the motor neurons would learn to move the paddle and intercept the ball. Any time the neurons missed an interception, they would be stimulated randomly, while successfully intercepting the ball meant they would receive predictable stimulation.
Why might random stimulation in response to error cause the neurons to learn to play the game? Kagan and his team used the idea behind the free energy principle, developed by Karl Friston, a collaborator on their paper, to inform their hypothesis. As Kagan explained, the free energy principle says that a system will minimize the surprise or uncertainty in its environment. “What we did was give the neuron feedback randomly if it got [the game] wrong. If the free energy principle is true, then the system should reorganize itself to minimize randomness,” Kagan said. This means that to minimize the amount of random stimulation they received, the in vitro neurons would need to learn to play the game.
This learning could be done in one of two ways: either the neurons could create a model or a “belief system” so that the network can respond and match the model with the real world, or they could physically act upon their environment to change their surroundings. Kagan and his colleagues showed that in vitro neurons learned to move the paddle to play Pong, and biological neurons can thus be adaptive. “We found that these neurons want to act in a way that can minimize unpredictability, [and] we can see this by them learning to play Pong,” Kagan said.
They also uncovered some interesting and unexpected findings. Kagan explained that one of the intriguing results of the paper was their data on information entropy, which is the amount of information conveyed in an event. “It was really exciting because it showed that the cultures were able to distinguish between internal and external noise,” Kagan said. Essentially, they showed that the neurons could determine the difference between information generated on their own and information from an outside source, highlighting the specificity with which the neurons can source the signals they receive. “[It] makes sense because I can distinguish between my thoughts and your words, so there must be a way to break that up. But to see that you’re getting one response for external noise and one response for internal noise was pretty exciting,” Kagan said.
This system, which Kagan and colleagues aptly termed “DishBrain,” sits at the interface of neurobiology and computational technology. Short-term benefits of the system are numerous: drug discovery, disease modeling, and building a basic understanding of how neurons create intelligence. “All general intelligence that we have ever seen is biological—from flies to cats to humans,” Kagan said. Still, there are limits to biological neural networks. “This does not mean that you end up with a human in a dish. What it means is that neurons are this biomimetic material that can adapt to new information, so can you use it as an information processor,” Kagan said. “It offers us an ethically responsible way to move forward.”
Though many questions remain, Kagan and his colleagues at Cortical Labs are looking forward to digging deeper into their work and making new, exciting findings. Now, they’re working on perfecting their research infrastructure, from creating new biological environments to advancing their technology. “We’re trying to improve what we call the wetware (the cells), the hardware, [and] the software. We’re starting to do some disease modeling and drug testing, and all of these options are super exciting,” Kagan said. While many questions remain, neurons are certainly firing at Cortical Labs to help uncover more answers. This exciting research is sure to produce more interesting data that will guide the field of neuroscience and biological intelligence in the future.