A rudimentary form of speech recognition has been conducted using clusters of human brain cells connected to a system. It is anticipated that these systems will consume notably less power compared to traditional silicon chips for AI-related tasks.
Feng Guo, based at Indiana University Bloomington, commented, “This serves as a proof of concept demonstrating our capability. However, we still have a considerable distance to cover.”
Under specific growth conditions, plant organisms can give rise to clusters of brain cells referred to as brain organoids. Guo asserts that these organoids bear a resemblance to miniature neurons.
Guo explains that these organoids, measuring a few millimeters in size and comprising up to 100 million nerve cells, undergo growth over a period of two to three months. This stands in contrast to the approximately 100 billion nerve cells present in the human brain.
The organoids are placed atop a microelectrode array, utilized for both detecting muscle cell firing responses and transmitting electronic signals to the organoids. The research team has coined the term “Brainoware” for this setup.
Reportedly, Guo’s team endeavored to tackle the Hénon map equations using this system, as detailed in a report by New Scientist from March.
In a speech recognition task, the organoids were trained to differentiate one person’s speech from 240 sound clips featuring eight individuals articulating Japanese vowel sounds. The organoids received patterns of signals arranged in spatial configurations corresponding to the sound clips.
Initially, the organoids’ accuracy in distinguishing speech was around 30 to 40 percent. However, following training sessions spanning two weeks, their precision improved to 70 to 80 percent.
Guo characterizes this process as “dynamic learning,” noting that no progress was observed when the organoids were exposed to a substance impeding the formation of new connections between brain cells.
The educational process primarily involved repetitive exposure to audio recordings, without any feedback provided to the organoids regarding their accuracy—a phenomenon termed uncontrolled learning in AI research.
Guo highlights two primary challenges faced by conventional AI: high energy consumption and the inherent limitations of silicon chips, including processing and data segregation.
Among the many teams exploring the potential of biocomputing utilizing living brain cells to address these challenges is Guo’s group. For instance, as reported by New Scientist in 2021, brain cells are being trained by an Australian company called Cortical Labs to play Pong.
Titouan Parcollet from the University of Cambridge, specializing in natural language processing, suggests that long-term utilization of biocomputing remains a viable option for enhancing conversational recognition tasks.
Parcollet cautions against assuming that emulating the brain is necessary to achieve the feats of current deep learning models. He emphasizes the superior performance of existing deep learning models on specific tasks compared to the brain.
Guo clarifies that their current work focuses solely on speaker identification rather than content analysis. He expresses reservations about the prospects of speech recognition based on their findings.
Despite the potential for enhancing Brainoware’s functionality, a significant limitation is the short lifespan of the organoids, typically surviving for only one to two weeks. Guo’s team is actively working on extending this duration.
Addressing these limitations is crucial, according to Guo, for leveraging the computing capabilities of organoids in AI applications.
