Researchers at the Massachusetts Institute of Technology have demonstrated that artificial neural networks can develop their own internal machinery for learning. This work, detailed in a paper published in the Proceedings of the National Academy of Sciences (PNAS), shows that networks can learn effectively by reorganizing themselves with only basic rules about how connections should change.
The study, conducted by scientists at MIT's McGovern Institute for Brain Research and the K. Lisa Yang Integrative Computational Neuroscience (ICoN) Center, focused on a network of randomly connected units. By providing simple instructions on when and how these connections should be modified, the researchers observed the network reshaping itself into an effective information-processing system. This self-organization contrasts with conventional approaches where developers meticulously design the learning architecture.
Qianli Liao, a postdoctoral fellow at the ICoN Center and co-first author of the paper, explained that traditional artificial intelligence training relies on engineers defining the rules for connection changes. While this method is powerful for achieving specific tasks, Liao and colleagues suggest that networks designed this way may offer limited insights into how biological brains learn. The team's self-organizing network, however, performed at a level comparable to conventionally designed networks in tasks such as pattern recognition and image classification.
The researchers argue that a key factor for an artificial neural network to be biologically plausible is its capacity for self-organization without direct developer guidance. One specific rule used in the network strengthens connections when connected units activate simultaneously, forming foundational structures for learning circuits. This emergent behavior mimics aspects of gradient descent, a common optimization algorithm in machine learning.
The study's findings provide a framework for investigating the formation of learning circuits. While the research does not claim that brains use this exact mechanism, it opens avenues for understanding biological learning and could potentially inform the design of future AI systems that better emulate the human brain's adaptability. The work aligns with broader efforts at MIT to understand neural self-organization, as seen in other research exploring how the brain uses spatial computing for cognitive tasks.
The team's approach offers a potential pathway to developing AI that learns more like biological systems, which could lead to more capable and human-like machine intelligence.