A new approach to machine learning uses a neural network based on the nematode nervous system
Artificial intelligence (AI) researchers are working on neural networks that attempt to mimic how the human brain is organized, but even with rapid progress, neural networks are often in situ. It lacks the flexibility to change and adapt to unfamiliar situations. A neural network 'liquid' based on the nervous system of a ' nematode (nematode) ' with a long, filamentous body was developed to realize such adaptability, and has unprecedented speed and flexibility. reportedly showing.
Closed-form continuous-time neural networks | Nature Machine Intelligence
Researchers Discover a More Flexible Approach to Machine Learning
In 2020, a research team led by Ramin Hasani and Matthias Lechner at the Massachusetts Institute of Technology introduced a new type of neural network, the Liquid Neural Network (Liquid), inspired by small nematodes. Did. Liquid has made dramatic progress in 2022, and has become versatile enough to replace conventional networks in certain applications. According to University of California, Berkeley roboticist Ken Goldberg, experiments show that Liquid works faster and more accurately than 'continuous-time neural networks' that model systems that change over time. ``Liquid offers an elegant and compact alternative,'' Goldberg said.
Hasani and Lechner, who led the design of Liquid, said that ``C. elegans is an ideal organism for exploring ways to create adaptive neural networks that can flexibly adapt to new situations.'' I noticed many years ago. C. elegans is one of the few organisms whose nervous system has been fully mapped, and from its one-millimetre-long nervous system, various advanced behaviors such as locomotion, foraging, sleep, copulation, and even learning from experience are demonstrated. can take ``Nematodes live in a real world where changes are constantly occurring, and they can do well in any situation,'' Lechner explains why he focused on nematodes.
Liquid is a neural network that characterizes the state of the system at any moment by linking and interdependent neurons, so it is very different from conventional neural networks that can only obtain results at a specific moment. . Liquid also differs in how it handles synapses, which are connections between artificial neurons. In standard neural networks, the strength of a synaptic connection can be represented by a single number as a 'weight'. On the other hand, in Liquid, the exchange of signals between neurons is a stochastic process governed by a 'nonlinear' function, which means that it does not return a response proportional to the input.
While traditional neural network algorithms are fed a large amount of data and are set by adjusting the optimal values of the 'weights' during training, Liquid can change the underlying equations based on observed inputs. It is more adaptable. When we tested the operation of an autonomous car, conventional neural networks were only able to analyze visual data from the car's camera at regular intervals, but Liquid is a very small '19' by machine learning standards. Despite being composed of 253 neurons and 253 synapses, it seems that it was able to demonstrate higher responsiveness. ``The model can sample complex roads more frequently, for example winding roads,'' said Daniela Luz, co-author of the paper.
On the other hand, for non-linear equations that represent synapses and neurons, which normally require many calculations on a computer to find a solution, Liquid uses a small number of synapses and neurons because it performs calculations with software that is individually adapted to synapses and neurons. was running very slowly, Lechner said. However, in a new paper published in November 2022, the research team proposed a new network that avoids this shortcoming, eliminating the need to solve nonlinear equations with difficult calculations and providing nearly accurate approximate solutions that can be obtained with basic calculations. It shows that the form of finding greatly improves the processing speed by reducing computation time and energy.
Sayan Mitra, a computer scientist at the University of Illinois at Urbana-Champaign, said of Liquid, 'Their method outperforms competitors by 'several orders of magnitude' without sacrificing accuracy.' talking about greatness. Sriram Sankaranarayanan, a computer scientist at the University of Colorado at Boulder, said, 'The main contribution of their work is the stability and other superior properties built into these systems by their very structure. Complex enough for interesting things to happen, but not complex enough to lead to chaotic behavior.'
A group at MIT is testing the latest liquid with an autonomously flying drone. Initial testing has been done in forests, but it is hoped that in the future we will move to urban environments to see how it copes with new conditions. Mr. Hasani also talks about the prospects, ``Liquid neural networks can perform brain activity simulations on a scale that could not be realized before.''