Use living human brain cells to build an AI system! Speech recognition has been successful and unsupervised learning is possible

An AI system composed of a "mini brain" and microelectrodes built from real human brain cells has been able to perform speech recognition -

From hundreds of The kind that accurately identifies a specific person's voice in a sound clip.

Recently, a very cutting-edge brain-inspired research was published in a sub-journal of Nature

This special The AI ​​system can even do unsupervised learning: The researchers just play the audio clip over and over again without providing any form of feedback to tell the system whether the answer is right or wrong.

In the end, after two days of training, the accuracy of the system increased directly from the initial

51% to 78%. How is this achieved?

organoid neural network is coming

The main purpose of inventing this system is to solve the problems of high energy consumption of silicon chips.

Generally speaking, the solution to this problem usually relies on brain-inspired computing

However, most of the "traditional" brain-inspired chips designed based on this idea are directly based on digital electronic principles , it is indeed limited in its ability to completely imitate brain functions

Here, the research directly used something called

"organoid":It refers to mini-organs that can be grown in the laboratory using human stem cells, and contain some of the key characteristics of their representative organs.

Specifically, the researchers connected

brain organoids

(shaped like small balls) composed of living brain cells to high-density microelectrode arrays to construct A system called "Brainoware" was released.

The role of microelectrodes in Brainoware is to send electrical signals to organoids to achieve the purpose of transmitting information to the "brain"; secondly, to detect the discharge of brain nerve cells The response is then handed to the external device for reading and parsing.

This system can exhibit functions similar to neural networks and can perform unsupervised learning

By connecting it to specific hardware, it can be trained for speech recognition.

In the specific task, the researchers converted 240 audio clips of 8 people speaking Japanese vowels into signal sequences and sent them to the system so that the system could recognize each person's voice

Initially, Brainoware's accuracy was only 30%-40%

After two days of training, it was able to identify specific targets with 78% accuracy Speaker

The author emphasizes here that the so-called training is just repeating audio clips without giving any feedback, which is the so-called unsupervised learning.

However, it should be noted that currently Brainoware can only identify who is speaking, but cannot understand any speech content.

After the experiment was completed, the researchers tried to use drugs to block the formation of new connections between nerve cells in the brain

After the experiment, it was found that after using this method of operation, the accuracy of the system improved. There is no improvement. The author explains that this shows that Brainoware's learning ability depends on

neuroplasticity

.

Will the computers of the future be made of brains? In March this year, the team actually used this system to try to predict the Hénon diagram (a dynamical system that can exhibit chaotic behavior in the field of mathematics)

.

The results of Brainoware were also after 4 days of unsupervised learning

(each day represents a training cycle), and it was found that it can predict more accurately than artificial neural networks without long-term memory units.

In contrast, the former only experienced less than 50 training cycles

Going a little further, an Australian research team tried to educate the "brain on the plate" to play table tennis. Surprisingly, it learned it in just five minutes, 17 times faster than artificial intelligence.

So in the future, will computers be composed of brains?

This is not yet certain

As the author of this article introduces, their research is currently a

proof-of-concept

, and there are still many problems to be solved later:

For example, although the performance of the Brainoware system can be further improved, the biggest problem is that organoids can only survive for one to two months

Also, although Brainoware itself does not require much power consumption, the power consumption level of the external devices that maintain its operation is not low.

What needs to be rewritten is: A series of other questions

In general, some scientists predict that a truly universal biological computing system may take decades to create.

In any case, this research will help us further understand the mysteries of human brain learning and other issues

The above is the detailed content of Use living human brain cells to build an AI system! Speech recognition has been successful and unsupervised learning is possible. For more information, please follow other related articles on the PHP Chinese website!