Researchers at the Massachusetts Institute of Technology ( MIT ) are working on a new neural network with the ability to learn constantly. They call the network ” liquid ” because of this adaptability that could result in systems with algorithms that modify themselves as they perform their work to adapt to possible new conditions.
Currently, most neural networks go through a “training” phase, to later be delivered as a ready-made product, which does not change anymore. The idea of MIT’s new technology is to have a “liquid” neural network that would never fail to learn, making changes as it is exposed to new data.
Such a system has many potential advantages, especially in autonomous cars example. A neural network capable of adapting would not depend on constant updates from a central station and would always be improving its ability to drive in the context in which the car is actually used by its driver.
” This is a way forward for the future of robotics control, natural language processing, video processing. Any form of data processing over time. The potential is really significant. “
Ramin Hasani, lead author of the published study in the new neural network
Hasani also says that the idea for the “liquid” neural network developed at MIT came from a small worm just a millimeter long, Caenorhabditis elegans. With its tiny size, the nematode has only 302 neurons, but it still manages to generate complex and unexpected dynamics. The researcher says, then, that he was inspired by the way that C. elegans neurons communicate with each other to make a neural network with a similar functioning.