Use Your Brain

Machine learning algorithms aim to replicate certain aspects of the human brain's function, particularly in tasks related to learning and decision-making. Inspired by the brain's neural networks, artificial neural networks have become a prominent approach in machine learning. These algorithms attempt to simulate the interconnected neurons and synapses found in biological brains, with layers of nodes and weighted connections that enable them to recognize patterns, make predictions, and classify information.

However, it is important to recognize that these machine learning algorithms are imperfect approximations of the brain's intricate workings. Despite their progress, they fall short in several aspects. One significant challenge lies in energy efficiency. While the brain operates on remarkably low power consumption, many machine learning models, especially large deep learning models, demand substantial computational resources and energy consumption. This inefficiency poses environmental concerns and limits the scalability of AI systems.

Moreover, when dealing with highly complex information or tasks that require a nuanced understanding of context, machine learning algorithms often struggle compared to biological brains. The human brain excels in contextual understanding, generalization, and adaptability, whereas AI models may struggle with ambiguity or unexpected situations.

These limitations underscore the current roadblocks that are hindering further advancements in AI. As researchers push the boundaries of machine learning, innovations are crucial to address these challenges. For a team led by researchers at Indiana University Bloomington, the path forward involves the cooperation between biological and artificial neural networks. They have developed a proof of concept system called Brainoware that leverages both clusters of brain cells and traditional silicon-based computing to work together in solving problems. They have demonstrated that Brainoware is capable of tasks like speech recognition and performing mathematical operations.

The researchers created small clusters of brain cells, called organoids, by treating pluripotent stem cells such that they would differentiate into neurons. These cells were then interfaced with a computer via an array of electrodes that can both stimulate the cells with electrical current, and measure signals produced by the cells. Traditional machine learning algorithms run on the computer, and serve to interpret the signals produced by the organoid.

Using the Brainoware system, an experiment was conducted in which hundreds of audio clips, from eight individuals, were converted into electric signals and fed into the cells via the electrode array. The signals produced by the organoid in response to this stimulus were captured and forwarded into an artificial neural network that attempted to identify the speaker. It was found that Brainoware could decode these signals with an average accuracy rate of 78%. This is well below what an entirely artificial system can achieve, however, it is an important first step in proving the concept.

In another demonstration, Brainoware was shown to be capable of doing math, predicting a Henon map successfully. Despite these successes, there is much work yet to be done. But with future advances, this hybrid approach could transform machine learning, enabling such systems to solve far more complex problems than what is possible today. Furthermore, the major issues we see today with energy efficiency could be alleviated. Whereas an artificial neural network can easily consume millions of watts of energy in a day, a typical human brain only requires 20 watts of energy.

The team imagines that one day Brainoware could assist researchers in studying neurological diseases, like Alzheimer’s. There is also a tremendous potential in the system to help us understand many other secrets of the brain.

But there are other practical issues with using Brainoware aside from improving the accuracy. Keeping the cells alive and functional, for example, is a big challenge. And as the system scales up, that challenge will only grow. However, with some work, this innovation could put a powerful new tool in the hands of machine learning researchers.