Balls of human brain cells grown in a dish, known as organoids, have been linked to computers and used to solve mathematical equations. The work is an early step towards using living brain tissue as a form of artificial intelligence, but this goal may raise ethical questions in the future, researchers say.
In a paper posted online before peer review, at Indiana University Bloomington and his colleagues say they have created 鈥渓iving AI hardware that harnesses the computation power of 3D biological neural networks in a brain organoid鈥�. The paper states that 鈥淏rainoware鈥�, as the researchers call it, can learn from training data and that experiments show it could have real-world applications.
Computer-based AI is getting very good at certain tasks, but this progress is being achieved by creating ever bigger and more energy-intensive AI systems, and training them on ever bigger data sets. For instance, the AlphaGo system that first beat humans at the game Go was trained on 160,000 games, more than any person could play in a lifetime.
Advertisement
Human brains use only around 20 watts of power, and people need to play far fewer games to become good. So some researchers 鈥� using living brain cells as AIs.
In 2021, New 杏吧原创 revealed that at in Australia and his colleagues had trained flat layers of human and mouse brain cells to play the computer game Pong. The 鈥淒ishBrains鈥�, as the team called them, learned much faster than conventional AIs.
Guo鈥檚 team instead used three-dimensional structures known as brain organoids. If stem cells are grown in the right conditions, they spontaneously form brain organoids, which don鈥檛 grow larger than a few millimetres wide as they lack blood vessels.
The researchers used human brain organoids to solve a non-linear equation called a H茅non map, which is difficult to predict because of its chaotic behaviour. The paper states that Brainoware outperformed conventional AIs without a so-called long short-term memory (LSTM) unit, but was less accurate than AIs with a LSTM.
However, there are few details in the paper and Guo didn鈥檛 respond to New 杏吧原创.
It is an interesting approach but doesn鈥檛 demonstrate real-world applications, says at the University of Edinburgh, UK, who has . 鈥淭he prediction is not incredibly impressive,鈥� he says.
And while the paper says that Brainoware learned over time, it isn鈥檛 clear how feedback was provided, Lellep says.
鈥淎lthough their preprint does need more details, it is an exciting idea to explore,鈥� says Kagan. 鈥淥rganoids are an exciting next step in using biological neurons for information processing, something we鈥檝e also been exploring for the last 15 months with various collaborators, and they can show a lot of interesting patterns of activity.鈥�
Brain organoids are very small and disorganised compared with real brains, and their computational capabilities are pretty rudimentary, says at the MRC Laboratory of Molecular Biology in Cambridge. 鈥淚 think it鈥檚 probably a stretch to compare brain organoids with current AI,鈥� she says.
Researchers are trying to create more advanced organoids, for instance by creating some kind of circulatory system so they can grow larger. 鈥淲hether these push them beyond an ethical limit is something we certainly want to avoid, and the scientific and ethics community is coming together to define where that limit would be, before we reach it,鈥� says Lancaster.
鈥淎t this stage, we are probably many years away from reaching real ethical boundaries, but we want the discussion now and not when there is a conflict of interest,鈥� says at Johns Hopkins University in Baltimore, Maryland, whose team is providing Kagan with brain organoids that should have better long-term memories.
bioRxiv
