Development of the first 'liquid' neural network, a machine learning system that not only updates in the training phase, but continues to learn from experience.
Created by the group of the Massachisetts Institute of Technology (MIT) in Boston coordinated by Ramin Hasani, it is presented in an article published on the ArXiv website, which collects research pending review by the scientific community.
The result, the research authors note, promises important applications in artificial intelligence systems of self-driving vehicles and related to medical diagnoses, based on data that change continuously over time.
Hasani and his team were inspired by the nerve networks of living things, particularly those of the tiny worm Caenorhabditis elegans, used as a model in biology studies.
"It is an organism that has only 302 neurons, yet - explains Hasani - it can generate unexpectedly complex dynamics".
The secret of these neural networks, the expert continues, are flexible algorithms, capable of continuously modifying their equations to adapt them to the new data collected from time to time: "a way forward - he observes - for the future of robot control ".