Neurons Receiving Input From a Computer Chip Demonstrate Learning in a Game of “Pong”
Post by Lani Cupo
The takeaway
The future of artificial intelligence is reimagined with DishBrain, a new technology that combines neurons, grown in a petri dish, with stimuli and feedback from electrical circuits. The authors present a “Pong”-playing system that they claim embodies the foundation of sentience.
What's the science?
Neurons and computer hardware share a common language: electricity. It has been theorized that combining real biological neurons with silicone hardware would allow a synthetic system to represent complexity that modern computers (using binary) alone cannot capture. Previously, however, the advantages of biological neural circuits have not been integrated into digital, silicone systems. Recently in Neuron, Kagan and colleagues present a form of synthetic biological intelligence called DishBrain: a petri dish of neurons embedded with a multi-electrode array that provides electrophysiological stimulation and recording to create a simulated game world that allows the neural circuit to learn the game “Pong” in real-time.
How did they do it?
The authors first generated cortical neurons from two sources: human pluripotent stem cells and mouse embryos. These neurons were then allowed to mature on plates made of multi-electrode arrays, developing into complex, interconnected neural circuits. The authors developed a system that leveraged software to not only record electrical activity from the neurons but also provide noninvasive electrical stimuli that mimic action potentials. In order to demonstrate real-time learning, the authors simulated the game “Pong”. They provided “sensory” information representing a ball moving on a trajectory. Electrophysiological activity in a predefined “motor” cluster of neurons was recorded, representing moving a paddle up or down. If the movement would result in an interruption of the ball’s trajectory, a predictable stimulus was presented to all the neurons at once, serving as positive feedback. If, however, the movement would not interrupt the ball’s path, an unpredictable stimulus was provided. The authors examined the hit-miss ratio, which they defined as the average rally length, as a metric of learning. As experimental controls, they compared the cell cultures to three different conditions: 1) Petri dishes with only cell culture media (no cells), 2) rest sessions, where the cells could control the paddle, but received no "sensory input", and 3) sessions that replicated the experiment, but where the paddle was controlled by random noise. With the recordings of cell firings, the authors examined functional connectivity within the culture and how cell activity related to performance in the task.
What did they find?
At the start of the sessions, they found that human stem cells performed worse than the mouse neurons and controls at the task, possibly due to increased exploratory behavior. However, by the end of the session, both mouse and human-derived cultures performed better than controls, and the human cultures performed slightly better than the mouse, indicative that learning occurred over the session. Importantly, the authors found that feedback was necessary for learning - the cultures only improved in average rally length when feedback was provided to the system. Overall, connectivity within the culture was stronger during gameplay than during rest, which could suggest a direct relationship between activity in “sensory” regions receiving input and “motor” regions supplying output. Increased neuron firing was also related to better performance in the game, although most important was that neuron activity was fairly symmetrical across the surface of the actual physical chip, suggesting a balanced culture.
What's the impact?
Because of the learning in response to environmental stimuli, the authors present DishBrain as a sentient form of synthetic biological intelligence, as the system was “responsive to sensory impressions through adaptive internal processes.” While advances would still be required in terms of hardware and “wetware” of the biological interface, the authors suggest a new direction for artificial intelligence. Their findings not only manifest what was previously confined to the realm of science fiction in terms of inventions but also some of the moral and ethical quandaries as well.
