800,000 neurons in a bowl learned to play pong in just five minutes

Scientists have just taught hundreds of thousands of neurons in a bowl to play pong. Using a series of strategically timed and placed electrical zaps, the neurons not only learned how to play in a virtual environment, but played better over time – with longer rallies and fewer misses – and exhibited a level of adaptation previously thought impossible became.

Why? Imagine literally taking a piece of brain tissue, digesting it into individual neurons and other brain cells, tossing them (gently) onto a plate, and now outside of a living host, teaching them to respond to a new task and adapt on their own with electric zaps.

It’s not all fun and games. The biological neural network joins its artificial cousin, DeepMind’s deep learning algorithms, in a growing pantheon of attempts to deconstruct, reconstruct, and one day dominate some sort of general “intelligence” based on the human brain.

The brainchild of the Australian company Cortical Labs, the whole setup is dubbed DishBrainis the “first real-time synthetic biological intelligence platform,” according to the authors of a paper published in this month neuron. The setup, smaller than a dessert plate, is extremely elegant. It connects isolated neurons to chips that can both record the cells’ electrical activity and trigger precise zaps to alter that activity. Similar to brain-machine interfaces, the chips are controlled with sophisticated computer programs without human intervention.

The chips act as a bridge for neurons to connect to a virtual world. As translators for neural activity, they can fuse biological electrical data with silicon bits, allowing neurons to respond to a digital game world.

DishBrain is designed to be expanded to more games and tests. Because the neurons can sense their environment, adapt to it, and output their results to a computer, they could be used as part of drug screening tests. They could also help neuroscientists better decode how the brain organizes its activity and learns, and inspire new methods of machine learning.

But the ultimate goal, explained Dr. Brett Kagan, Chief Scientific Officer at Cortical Labs, is to help harness the inherent intelligence of living neurons for their superior computing power and low energy consumption. In other words, why not just use the original compared to neuromorphic hardware that mimics neural computations?

“Theoretical generalized SBI [synthetic biological intelligence] may come ahead of artificial general intelligence (AGI) because of the inherent efficiency and evolutionary advantage of biological systems,” the authors write in their article.

To meet DishBrain

That DishBrain The project started with a simple idea: neurons are incredibly intelligent and adaptable computing machines. Recent studies suggest that each neuron is a supercomputer in its own right, with branches once thought to be passive functioning as independent minicomputers. Like people within a community, neurons also have the ability to connect to different neural networks that change dynamically with their environment.

This level of parallel, low-power computation has long inspired neuromorphic chips and machine learning algorithms to mimic the brain’s natural abilities. While both have made advances, neither has been able to replicate the complexity of a biological neural network.

“From worms to flies to humans, neurons are the starting point for general intelligence. So the question was, can we interact with neurons to use this inherent intelligence?” Kagan said.

Enter DishBrain. Despite their name, the plated neurons and other brain cells are from an actual conscious brain. As for “intelligence,” the authors define it as the ability to gather information, gather the data, and adjust firing activity — that is, how neurons process the data — in a way that helps adapt to a goal; For example, learn to quickly put your hand on the handle of a boiling hot pan without burning the edge.

The setup, true to its name, begins with a dish. The underside of each is covered with a computer chip, HD-MEA, which can record stimulated electrical signals. Cells are then placed on top of this, which are either isolated from the cortex of mouse embryos or stem from human cells. The dish is immersed in a nourishing liquid to allow the neurons to grow and thrive. As they mature, they grow from wobbly blobs to scrawny forms with vast networks of twisting, intertwined branches.

Within two weeks, the neurons of mice in their tiny houses self-organized into networks and were bursting with spontaneous activity. Neurons of human origin – skin cells or other brain cells – took a little longer, building networks in about a month or two.

Then came the training. Each chip was controlled by commercially available software that connected it to a computer interface. Using the system to stimulate neurons is similar to providing sensory data – like what comes from your eyes when you focus on a moving ball. Recording the neurons’ activity is the result — that is, how they would react (if they were in a body) to you moving your hand to hit the ball. DishBrain was designed to integrate the two parts in real time: similar to humans playing pong, the neurons could theoretically learn from past failures and adapt their behavior to hit the virtual “ball”.

ready player DishBrain

That’s how Pong works. A ball quickly bounces across the screen and the player can slide a tiny vertical paddle – which looks like a bold line – up and down. Here the “ball” is represented by electrical zaps based on its position on the screen. This essentially translates visual information into electrical data that the biological neural network can process.

The authors then defined different regions of the chip for “sensation” and “movements”. For example, a region captures incoming virtual ball motion data. One part of the “motor region” then controls the virtual paddle up, another down. These assignments were arbitrary, the authors explained, meaning the neurons inside had to adjust their firings to shine in a match.

So how do they learn? When the neurons “hit” the ball, that is, showed the appropriate type of electrical activity, the team then tapped them at the same frequency at that spot. It’s a bit like imparting a “habit” to the neurons. If they missed the ball, they were zapped with electrical noise that disrupted the neural network.

The strategy is based on a learning theory called the free energy principle, Kagan explained. The basic assumption is that neurons hold “beliefs” about their environment, and adjust and repeat their electrical activity so that they can better predict the environment, either by changing their “beliefs” or by changing their behavior.

The theory worked. In just five minutes, both human and mouse neurons rapidly improved their gameplay, including better rallies, fewer aces – where the bat didn’t intercept the ball without a single hit – and long gameplays of more than three consecutive hits. Surprisingly, mouse neurons learned faster, although they were eventually outperformed by humans.

The stimulations were critical to their learning. Separate attempts with DishBrain without electrical feedback performed far worse.

Continue to play

The study is proof of concept that neurons in a shell can be a sophisticated learning machine and even show signs of sentience and intelligence, Kagan said. This does not mean that they are conscious – rather they have the ability to adapt to a target when “embodied” in a virtual environment.

Cortical Labs isn’t the first company to test the limits of the computing power of isolated neurons. Already in 2008, Dr. Steve Potter of the Georgia Institute of Technology and his team found that with just a few dozen electrodes, they could stimulate rat neurons to show signs of learning in a dish.

DishBrain has a head start with thousands of electrodes compacted into each build, and the company hopes to use its biological power to aid drug development. The system, or its future derivatives, could potentially serve as a surrogate for the microbrain to test neurological drugs or gain insight into the neurocomputational abilities of different species or brain regions.

But the long-term vision is a “living” bio-silicon computer hybrid. “Integrating neurons into digital systems can enable performance that would not be achievable with silicon alone,” the authors write. Kagan envisions creating “biological processing units” that weave the best of both worlds together for more efficient calculations – while shedding light on the inner workings of our own minds.

“This is the beginning of a new frontier in understanding intelligence,” Kagan said. “It not only touches on the fundamental aspects of what it means to be human, but also what it means to even be alive and intelligent, to process information and to be sentient in an ever-changing, dynamic world.”

Photo credit: Cortical Labs