Artificial intelligence is making significant strides in areas like advanced mathematics, complex reasoning, and personal computing. However, today’s AI algorithms could still gain insight from tiny organisms such as microscopic worms.
Liquid AI, a startup born from research at MIT, is set to unveil several innovative AI models that utilize a unique type of “liquid” neural network. These new models promise to be more efficient, require less power, and offer greater transparency compared to traditional networks that support various applications like chatbots, image generation, and facial recognition.
Among the new models introduced are tools designed for detecting fraudulent financial transactions, managing self-driving cars, and analyzing genetic information. Liquid AI is licensing these technologies to external companies and showcased them at an event at MIT. The company has attracted investments from notable backers, including Samsung and Shopify, both of which are currently experimenting with Liquid AI’s technologies.
Ramin Hasani, co-founder and CEO of Liquid AI, expressed excitement about the company’s growth. Hasani, who co-developed the concept of liquid networks while he was a graduate student at MIT, was inspired by the microscopic worm C. elegans. This tiny worm, often found in soil or decaying plant matter, has had its nervous system fully mapped and is capable of intricate behaviors despite having only a few hundred neurons. “What began as a scientific project has now evolved into a fully commercialized technology poised to deliver value for businesses,” Hasani stated.
In a traditional neural network, the attributes of each simulated neuron are determined by fixed values known as “weights” that influence their activation. In contrast, a liquid neural network operates with equations that predict neuron behavior over time, allowing for the resolution of interconnected equations as the network operates. This dynamic structure enhances the network’s efficiency and adaptability, permitting continued learning even after the initial training phase. Moreover, liquid neural networks offer transparency, as their outputs can be traced back to see how they were produced.
In 2020, researchers demonstrated that a liquid neural network containing just 19 neurons and 253 synapses—remarkably small by today’s standards—could effectively manage a simulated self-driving car. Unlike conventional neural networks that analyze visual data at discrete intervals, liquid networks efficiently capture the temporal changes in visual information. In 2022, the founders of Liquid AI developed a method that streamlined the mathematical requirements for utilizing liquid neural networks, making them practical for real-world applications.
.
