Brain activity and AI systems

When human programmers analyze and mentally comprehend code, their brains trigger very specific patterns of activity. These patterns reveal which characteristics the code under consideration has and are surprisingly similar to those generated by artificial neuroyal networks during the same task. That's what researchers from the Massachusetts Institute of Technology (MIT) found.

Using functional magnetic resonance imaging (fMRI), they found that reading program code is significantly different from how people read normal text. When looking at and analyzing code, not only the visual cortex and the language center were active. The so-called multiple demand network (MDN) also showed significant activity. This is associated with complex tasks such as logical thinking and problem solving.

The researchers Liesen the subjects 72 Python scripts read and analyzed. They had grouped the code into four different categories of tasks. In one category, for example, the control flow of a program had to be understood and the outcome predicted. In another experiment, the researchers had different artificial neural networks read and evaluate the same program codes as the human subjects. This revealed that humans and machines produce quite similar patterns of activity and neural representation. The encoding in the MD system of the human brain showed particularly high agreement with the artificial networks.

The researchers presented their paper at the Neural Information Processing Systems conference. Nevertheless, they were able to identify one major difference: in contrast to human subjects, the AI used only the language center. Humans process program code in a more complex way. The researchers hope that their findings will provide new food for thought for the development of source code-comprehending AI.