Computation: neural networks that can reverse engineer data and create algorithms | Computational natural sciences

A research paper discusses a method that uses neural networks to extract observed data and create a simple algorithm that a human can understand.Computational natural sciencesPublished in

Data-driven algorithmic discovery is a method that allows the discovery of the underlying logic and rules behind experimental data sets. Using this technique, researchers may be able to generate new insights hidden in high-dimensional data. However, typical approaches to algorithm discovery are often incomprehensible to humans, rely on learning code written by other programmers, or are not scalable due to the large search space for potential functions.

Now, Milo Li and his colleagues propose a way to write understandable and executable computer code from data using deep learning algorithms inspired by neurobiology. This algorithmic design uses a symbolic (and therefore human-readable) approach to training rather than the typical gradient-based approach, so neural networks are inherently interpretable and able to discover logical rules from data. It can be used for. These results can then be translated into pseudocode or executable programming languages.

Lee and others found that the method they proposed works very similarly to a human-designed algorithm, and that the algorithm used to detect shapes in images is similar to a human-designed algorithm. We have shown that it is possible to discover algorithms that outperform In particular, based on repeated observations of Conway's Game of Life (the famous code that develops patterns on a 2D grid based on raw inputs), we have been able to discover executable code that mimics it. This indicates that this method has the potential to understand complex self-organization patterns.

Although this method is best suited to problems with algorithmic solutions, Lin and colleagues say a potential application that their proposed method could investigate further is the relationship between DNA base sequences, structure and function, lists discoveries and so on.

doi:10.1038/s43588-024-00593-9

“Highlights from Nature Magazines” is a translation of a statement prepared for the press by Nature's public relations department. If you require more accurate and detailed information, please be sure to refer to the original paper.

Return to the list of “Highlights” articles.