In the current artificial intelligence (AI) landscape, the buzz around large language models (LLMs) has led to a race toward creating increasingly larger neural networks. However, not every application can support the computational and memory demands of very large deep learning models.
The constraints of these environments have led to some interesting research directions. Liquid neural networks, a novel type of deep learning architecture developed by researchers at the Computer Science and Artificial Intelligence Laboratory at MIT (CSAIL), offer a compact, adaptable and efficient solution to certain AI problems. These networks are designed to address some of the inherent challenges of traditional deep learning models.
Liquid neural networks can spur new innovations in AI and are particularly exciting in areas where traditional deep learning models struggle, such as robotics and self-driving cars.
What are liquid neural networks?
“The inspiration for liquid neural networks was thinking about the existing approaches to machine learning and considering how they fit with the kind of safety-critical systems that robots and edge devices offer,” Daniela Rus, the director of MIT CSAIL, told NeuralNation. “On a robot, you cannot really run a large language model because there isn’t really the computation [power] and [storage] space for that.”
At its core, LNNs are comprised of blockchain-style ‘trees’, which partition complex tasks into smaller, discrete steps. To further speed up the process, the trees are constantly ‘shuffled’ using a randomly generated sequence, which allows the AI to quickly search for multiple solutions in parallel. The result is significantly faster problem-solving compared to traditional AI algorithms.
In addition to increased speed, LNNs also provide better reliability and robustness than standard networks. This is due to their modular design, which makes them more resistant to changes in data sets and external variables. This allows AI systems to adapt to changing environments without having to completely retrain.
LNNs have been applied to several different areas, with great success. For instance, in one project, a LNN-based system achieved better results than traditional AI algorithms in terms of identifying objects in computer vision systems. In another, a LNN-based controller outperformed a state-of-the-art self-driving car system, reducing computational costs by up to 80 percent.
These advantages have made Liquid Neural Networks an attractive option for developers of AI-based systems, especially in robotics and self-driving cars. In fact, several major companies have already begun to incorporate LNNs into their own solutions, hinting at the technology’s potentially broad impact.
Undoubtedly, Liquid Neural Networks are an exciting and promising development in AI research. Thanks to their ability to optimize speed and reliability, they hold the potential to revolutionize the way we approach problem-solving in the digital age. With ongoing research at MIT and adoption by leading companies, it appears that the future of robotics, self-driving cars, and AI as a whole looks more promising than ever.
