A very basic kind of speech recognition has been achieved using computerized human brain cell balls. It is hoped that these systems will require a lot less energy than silicon chips for AI tasks. According to Feng Guo of Indiana University Bloomington, this is merely proof-of-concept to demonstrate their capability. They certainly still have a ways to go.
Brain organoids are clumps of nerve cells that develop from stem cells cultured under specific circumstances. Guo describes them as “like mini-brains.”
The organoids, which are a few millimeters wide and comprise as many as 100 million nerve cells, take two or three months to grow, according to him. There are about 100 billion nerve cells in the human brain.
The microelectrode array is then positioned on top of the organoids, serving as a means of both electrically stimulating the organoid and monitoring the firing of nerve cells in response. The system is dubbed “Brainoware” by the group.
Guo’s group attempted to solve equations known as a Hénon map using this system, according to a March New Scientist article.
In order to complete the speech recognition task, the organoids needed to be able to identify one person’s voice from 240 audio clips featuring eight people speaking Japanese vowel sounds. The clips were transmitted as signal sequences arranged in spatial patterns to the organoids.
According to Guo, the organoids’ first answers were between thirty and forty percent accurate. Following two days of training, they increased their accuracy to seventy to eighty percent.
He says “we call this adaptive learning”. There was no improvement if the organoids were exposed to a medication that prevented new connections from forming between nerve cells.
Guo claims that the training consisted only of having the organoids repeat the audio clips without any kind of feedback to let them know if they were correct or incorrect. In the field of AI research, this is referred to as unsupervised learning.
Guo identifies two major issues with conventional AI. Its high energy consumption is one. The other is the intrinsic limitations of silicon chips, which include the division of processing and information.
Several teams, including Guo’s team, are investigating whether biocomputing with living nerve cells can help address these issues. For example, according to a 2021 article in New Scientist, brain cells are being taught how to play Pong by an Australian company named Cortical Labs.
Is using intricate mini-brains to create artificial intelligence morally acceptable?
In the long run, biocomputing may play a role, according to Titouan Parcollet of the University of Cambridge, who specializes in conventional speech recognition.
Parcollet adds that it might be erroneous to believe that deep learning can only accomplish these kinds of tasks in the absence of something akin to the brain. “Current deep-learning models are actually much better than any brain on specific and targeted tasks.”
According to Guo, the work of him and his team is so streamlined that it only involves identifying the speaker—not the content of the speech. From the standpoint of speech recognition, the results don’t seem all that promising.
Guo claims that even if Brainoware’s performance could be enhanced, the fact that the organoids could only be kept alive for a month or two is still a significant problem. Extending this is what his team is working on.
He says that addressing those limitations is crucial if they hope to use the computational power of organoids for AI computing.
