Member-only story
Is “Liquid” ML the answer to autonomous driving?
How free-living, transparent nematodes provide inspiration for the next steps in Deep Learning
Caenorhabditis elegans, or C. elegans, is a free-living, transparent nematode. A roundworm, about 1mm in length that lives in temperate soil environments. Its nervous system consists of 302 neurons, yet it can generate surprisingly complex and diverse patterns.
On the other hand, we have Recurrent Neural Networks with millions of parameters and thousands of nodes. Still, their behavior stays fixed after the training phase. Consequently, their adaptability to changing environments is limited.
A network’s behavior remains fixed after the training phase. What happens, then, when the incoming data’s characteristics change during inference? How could we address this issue?
To this end, a new study by Ramin Hasani et al., a postdoc in the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), examine the properties of a new class of time-continuous recurrent neural network models. “Liquid” NNs can vary their equations’ parameters and adapt to the changing dynamics of time-series datasets.