Liquid AI Redesigns Neural Networks

Artificial intelligence may now be solving advanced math, using complex reasoning, and using personal computers, but today’s algorithms can still learn a thing or two from tiny worms.
Liquid AI, a startup that came out of MIT, today will unveil several new AI models based on a new type of “liquid” neural network that is more efficient, less power hungry, and more transparent than the ones it supports. everything from chatbots to image generators to facial recognition systems.
The new Liquid AI models include one to detect fraud in financial transactions, another to control self-driving cars, and a third to analyze genetic data. The company applauded the new models, which it licenses to third parties, at an event at MIT today. The company has received funding from investors including Samsung and Shopify, both of which are testing its technology.
“We’re growing,” said Ramin Hasani, founder and CEO of Liquid AI, who founded liquid networks as a graduate student at MIT. Hasani’s research was inspired by C. elegansmillimeter-long worm usually found in soil or decaying plants. A worm is one of the few creatures whose entire nervous system has been mapped, and is capable of incredibly complex operations despite having only a few hundred neurons. “It was once a scientific project, but this technology is fully commercialized and fully ready to bring profit to businesses,” Hasani said.
Within a typical neural network, each neuron’s simulated properties are defined by a fixed value or “weight” that affects its firing. Within a fluid neural network, the behavior of each neuron is governed by an equation that predicts its behavior over time, and the network solves a number of connected equations as the network operates. The design makes the network efficient and flexible, allowing it to learn even after training, unlike a conventional neural network. Liquid neural networks are also open to testing in a way that existing models are not, because their behavior can be re-evaluated to see how they generated the output.
In 2020, researchers showed that such a network with only 19 neurons and 253 synapses, very small by today’s standards, could control a simulated self-driving car. While a conventional neural network can only analyze visual data at static times, a fluid network captures how visual information changes over time very well. In 2022, the founders of Liquid AI found a shortcut that made the math work necessary for neural networks available for use.
Source link