Reservoir computing reveals computational capabilities of artificially cultured brains

The brain is a remarkable information processing system, consisting of billions of interconnected neurons that transmit and process information. To optimize its efficiency, the brain develops various modules responsible for different functions, such as perception and motor control. These modules, comprised of clusters of neurons, act as interconnected units that have remained relatively unchanged throughout evolution.

However, there are still many unresolved questions regarding how the brain’s network structure, including its modular organization, interacts with the physical and chemical properties of neurons to process information effectively.

Reservoir computing is a computational model inspired by the brain’s abilities. It employs a large number of interconnected nodes, forming a reservoir, which transforms input signals into more complex representations.

Recently, a research team from Tohoku University, led by Takuma Sumi, Hideaki Yamamoto, and Ayumi Hirano-Iwata, in collaboration with Yuichi Katori from the Future University Hakodate, developed a machine learning approach based on reservoir computing. They used this method to analyze the computational capabilities of an “artificially cultured brain” composed of rat cortical neurons derived from the cerebral cortex of rats.

Their findings, published in the Proceedings of the National Academy of Sciences on June 12, 2023, demonstrated the team’s progress. They began by recording the responses of the cultured neuronal network using techniques such as optogenetics and fluorescent calcium imaging. Subsequently, they decoded these responses using reservoir computing and discovered that the artificial cultured brain exhibited short-term memory lasting several hundred milliseconds. This memory could be utilized for classifying time-series data, such as spoken digits.

The reservoir computer based on biological neurons could be used to classify spoken digits even when the speakers were switched during training and testing. Classification accuracy after the switch decreased compared to when there was no speaker switching, but classification was achieved above chance level. Such classification was not possible when the input signal was directly decoded by a linear classifier, suggesting that biological neurons act as a generalization filter to improve the performance of reservoir computing. Credit: Yamamoto et al.

The research team also observed that samples with a greater degree of modularity demonstrated superior performance in classification tasks. Interestingly, they discovered that a model trained on one dataset could successfully classify another dataset within the same category. This indicates that the artificial cultured brain possessed the ability to filter information, enhancing the performance of reservoir computing.

Yamamoto further commented on the significance of these findings, stating that they contribute to our understanding of how information is processed within neuronal networks comprising biological neurons. Moreover, these findings bring us closer to the possibility of creating physical reservoir computers utilizing biological neurons.

Source: Tohoku University

Leave a Reply

Your email address will not be published. Required fields are marked *