MIT researchers demonstrated that artificial neural networks can build their own learning machinery without an external designer, according to a study published Oct. 9.
Self-Organizing Networks
Scientists at MIT's McGovern Institute for Brain Research and the K. Lisa Yang Integrative Computational Neuroscience Center published their findings in the journal PNAS. They showed that with simple instructions, a group of randomly connected units can reshape itself into an effective information-processing network.
Qianli Liao, a postdoctoral fellow at the center, noted that the brain does not have an external designer. The team designed their network to explore biological learning by letting units interact using basic rules.
Mimicking Biological Learning
In both the brain and the team's machine network, connected neurons that activate at the same time strengthen their connection. The researchers showed that a network governed by these rules can learn to recognize patterns and classify images.
Learning in this model is enabled by self-assembling motifs that reorganize using an optimization strategy similar to gradient descent. The team compared their network to systems using standard back-propagation and found it performed at a similar level.
Broader Implications
Tomaso Poggio, an MIT professor and McGovern Institute member, stated that finding a biological parallel to back-propagation could trigger significant progress in both artificial intelligence and neuroscience. Liao added that while self-organizing networks do not necessarily learn the exact same way brains do, they provide a critical starting point for biological plausibility.
Understanding these neural interactions could eventually inform the development of brain-computer interfaces or improve machine learning systems. However, the researchers emphasized that self-organization is currently the primary framework for measuring biological plausibility.
Key facts and where they come from
- MIT researchers showed that artificial neural networks can build their own learning machinery.
They show that it is possible for an artificial neural network to build its own learning machinery.
- The study was published in the journal PNAS on October 9.
In a paper published Oct. 9 in the journal PNAS
- The network begins with randomly connected units that follow basic rules to reshape themselves.
a group of randomly connected units can reshape itself into an effective information-processing network
