DishBrain, a neural network embedded in a computer chip, stunned the world in 2021 when it learned to play the popular video game Pong in just 5 minutes – a skill that allowed it to outperform some current artificial intelligence (AI) systems. Aiming to revolutionize AI with broad applications, the semi-biological chip received $407,000 in funding from a prestigious Australian military program.
DishBrain integrates around 800,000 neurons grown in a Petri dish in combination with an electronic silicon chip. It has demonstrated capabilities that would be difficult to achieve with electronic networks alone, even those that make up today’s best devices. The chip “brings together the fields of artificial intelligence and synthetic biology to create programmable biological computing platforms,” said one of the project leaders, Adeel Razi of University College London and Australia’s Monash University, in a statement.
The chip uses the computing power of biological cells and promises broad application perspectives such as robotics, advanced automation, brain-machine interfaces or the discovery of new drug molecules. It offers a significant strategic advantage for Australia, which could design an entirely new generation of machine learning systems and potentially develop the first truly autonomous devices — that is, ones that require little or no remote computing support. According to Razi, these devices require “a new breed of artificial intelligence that can learn throughout its lifecycle.”
:: THE T-SHIRT THAT SUPPORTS SCIENCE! ::
Outperform silicon devices
The computational superiority of biological cells has been extensively demonstrated with biomimetic materials that support neuromorphic computations. However, despite all attempts, no artificial network has been able to support the computational complexity of level 3, ie representing three different state variables at the same time. Although neural computations could be imaged in vivo, exploring them in vitro has many limitations.
To overcome these challenges, Australian researchers combined the innate intelligence of biological cells with the artificial materials that support standard computing systems. The combination creates a silicone soft tissue mixture. Among other things, DishBrain uses the adaptive calculation of neurons within a structured environment. “This new technological capability could eventually surpass the performance of existing purely silicon-based systems,” believes Razi.
Fluorescence microscopy shows neurons in the DishBrain chip. © Cortical Labs
Revolutionize artificial intelligence?
The biological neural network was developed from rodent embryonic cells or human induced pluripotent stem cells (hiPSC). The cells were cultivated in a high-density multi-electrode array and successfully demonstrated biological intelligence. In fact, the network of microelectrodes was able to read the activities of the cells and stimulate them with electrical signals.
Overall, the system takes advantage of the inherent property of neurons to share a “language” of electrical activity, allowing the silicon system to be connected to the network of living cells through electrophysiological stimulation and recording.
In an experiment last year, scientists tested the chip’s capabilities with the video game Pong. In this version of the game, neurons received a moving electrical stimulus that represented which side of the screen the ball was on and how far it was from the bat. Cells could then act on the paddle by moving it from left to right as in the game.
Figure summarizing the Pong gaming experience using the DishBrain chip. © Neuron
Would you like to remove ads from the site while still supporting us?
It’s easy, just subscribe!
At the moment, 20% discount on annual subscription!
Second, a basic reward system was established to assess how the semibiological network can reduce the unpredictability of its environment. If he manages to accept the ball with the bat, he receives a pleasant and predictable stimulus. On the other hand, if the ball misses, an uncomfortable and unpredictable four-second stimulus is triggered. In the experiment presented in the 2022 study, DishBrain adapted quickly and learned to play correctly in just 5 minutes! This reward system is critical because if stimuli are induced without feedback, the device will show no learning.
These results suggest that, similar to neurons in our brain, DishBrain exhibits sensitivity through adaptive processes. This suggests that the device has a learning ability throughout its lifetime, allowing it to acquire new skills without forgetting old ones. Namely, while current AIs can evolve by acquiring new skills, they are still limited in that they can “forget” certain skills or sub-skills. Hence the interest in semi-biological systems that will be able to adapt to change and apply previously acquired knowledge to new tasks, provided that their computing capacity, memory and the energy that feeds them allow it.
With the new funding, DishBrain’s designers can expand its capacity and assess how far the device can go and whether it can replace some current silicon-based devices. In addition, “we will use this grant to develop better AI systems that emulate the learning ability of these biological neural networks,” Razi concludes.