Scientists have fused brain-like tissue with electronics to make an ‘organoid neural network’ that can recognise voices and solve a complex mathematical problem. Their invention extends neuromorphic computing – the practice of modelling computers after the human brain – to a new level by directly including brain tissue in a computer.
The system was developed by a team of researchers from Indiana University, Bloomington; the University of Cincinnati and Cincinnati Children’s Hospital Medical Centre, Cincinnati; and the University of Florida, Gainesville. Their findings were published on December 11.
The study marks a significant advance in multiple areas of science and engineering. “It opens possibilities at the intersection of tissue engineering, electrophysiology, and neural computation,” Thomas Hartung, a professor at Johns Hopkins University, in the U.S., said.
The work comes against the backdrop of the staggering rise of artificial intelligence (AI), itself founded on the development of artificial neural networks – brain-like networks of neurons except they’re made with silicon chips – that can process large datasets that conventional computers struggle with.
The memory and processing separation
The hardware on which these neural networks run has a problem, however: the memory units and the data processing units are separate. When a neural network operates, the network will have to access the data in the memory unit, bring it over to the processing unit, and work on it – and it needs to do this many times over. If the problem is more complex, the time and energy demands increase further because the system will have to go back and forth between these units even more.
Scientists have tried to build more efficient neuromorphic chips that include some short-term memory, so they can avoid going back and forth just a bit. These chips have been used for applications like computer vision and speech recognition. But they can “only partially mimic brain functions, and there is a need to improve their processing capability and accounting for real-life uncertainty and improving energy efficiency,” the authors of the new paper, published in Nature Electronics, wrote.
So scientists are now considering using a biological neural network: a network of live brain cells. As the authors described in the paper, the brain spends only 20 W to do the same amount of work that AI hardware would use about 8 MW to drive artificial neural networks made of silicon chips. This difference by a factor of 400,000 is because, unlike AI hardware, brain cells store memory and process data without physically separating the two.
An ‘organoid neural network’
This new area of research, called biocomputing, uses biological components to perform computational processes. Last year, for example, a group of researchers from Australia cultured brain cells and trained them to play pong (a table-tennis-like videogame), in the process demonstrating the initial steps of long-term training. Their paper was published in the journal Neuron.
In the new study, the U.S. researchers used actual brain organoids to make an ‘organoid neural network’ and tested it to recognise speech and solve a complex mathematical problem.
Brain organoids are three-dimensional aggregates of brain cells. The scientists made them by extracting human pluripotent stem cells, which are cells that can develop to become almost any kind of cell within the human body, and made them into brain cells. Brain organoids that are aggregates of such cells have a mix of the different types of cells in the brain. In this case, they were neuron progenitor cells, early-stage neurons, mature neurons, and astrocytes (cells that maintain and protect neurons).
A three-layered computer
The team connected the brain organoid to an array of microelectrodes to form an organoid neural network, a type of artificial neural network containing a live organoid. The team then built this network into a system called a reservoir computer.
This machine contains three ‘layers’: input, reservoir, and output.
The input signals are routed to a reservoir, which is a black box – meaning its internal working can’t be tracked – whose purpose is to convert the signals into mathematical entities that the computer can ‘work on’ to find solutions. The output is a simple readout from the reservoir.
In this system, which the team calls ‘Brainoware’, the reservoir was the organoid neural network. It received inputs from the input layer in the form of electrical stimulation. The output layer was ‘normal’ computer hardware that had been modified to recognise Brainoware’s neural activity.
The researchers demonstrated Brainoware’s abilities by predicting a Henon map – a mathematical function that draws a curve on a graph that can be chaotic or not depending on the values of two variables. Brainoware could also tell which Japanese vowel an individual was voicing after ‘learning’ from 240 audio clips from eight speakers, over just two days. Its accuracy in the latter task was 78%, and without any external feedback on whether its inferences while learning were right or wrong.
‘Foundational insights’
More importantly, Brainoware was more accurate than artificial neural networks that lacked a short-term memory unit but slightly less accurate than those with one. But it achieved this comparable accuracy with less than a tenth of the training the artificial neural networks required. For example, to be able to predict a Henon map, Brainoware went through four epochs of training but the artificial neural network went through 50 epochs.
In a commentary published alongside the study, Lena Smirnova, Brian Caffo, and Erik C. Johnson, all professors at Johns Hopkins University, wrote, “It may be decades before general biocomputing systems can be created, but this research is likely to generate foundational insights into the mechanisms of learning, neural development, and the cognitive implications of neurodegenerative diseases.”
Their system still has some limitations that the researchers have also acknowledged, but it’s a start. Running Brainoware requires technical expertise and infrastructure to maintain a biological neural network. Organoids form a relatively heterogeneous mix of cell types, so not all organoids function the same way. But efforts are underway to achieve more uniform cell mixes.
Are organoids conscious?
Further, although this was the first system to use a three-dimensional culture of brain cells, it interfaced with the input layer only along one surface of the organoid. According to Dr. Hartung, who’s working on using organoids and AI instead of animals to test new drugs, the next steps could include optimising methods to encode input, improving the viability or maintaining uniformity of organoids in longer experiments, and tackling more complex computing problems.
Brainoware-like systems also confront us with ethical concerns. For example, Julian Kinderlerer, an emeritus professor at Delft University of Technology, the Netherlands, asked in a March 2023 article whether an organoid would have the same dignity as the donor of its cells. He also considered whether it would be fair to “use organoids in a mechanistic way without … being aware of their state of consciousness”.
Dr. Hartung nonetheless described the study as an “innovative and exciting proof-of-concept study of organoid intelligence, showing that brain organoids could be harnessed for adaptive reservoir computing.”
Joel P. Joseph is a freelance science journalist and researcher.
