FOLLOW KIM BELLARD
Chances are, you’ve read about AI recently. Maybe you’ve actually tried DALL-E or ChatGPTmaybe even GPT-4. Perhaps you can use the term Large language model (LLM) with some degree of confidence. But chances are you haven’t heard of “liquid neural networks” and didn’t get the worm reference above.
That’s the thing about artificial intelligence: it grows faster than us. Anything you think you know may be outdated.
Liquid neural networks are introduced for the first time in 2020. “We introduce a new type of time-continuous recurrent neural network model,” the authors wrote. They built the network based on the brain of a small roundworm, Caenorhabd elegant inflammation. The goal is networks that are more resilient, can change “on the fly” and will adapt to unfamiliar circumstances.
Researchers at MIT’s CSAIL have shown some remarkable progress. ONE new paper in robotics science discussed how they created “powerful flight navigation agents” using liquid neural networks to control drones autonomously. They state that these networks are “causal and adaptive to changing conditions” and that “their tests show a level of robustness in this decision-making that is unique to liquid networks.” .
ONE MIT . Press Release note: “deep learning systems have difficulty capturing causation, often overfit their training data, and fail to adapt to new environments or changing conditions…Unlike networks Traditional neural networks only learn during the training phase, the parameters of liquid neural networks can change over time, making them not only interpretable but also more resilient to unwanted data. or noisy.”
“We wanted to model the dynamics of neurons, how they function, how they release information, neuron to neuron, Ramin Hasania research affiliate at MIT and one of the co-authors, told Popular Science.
Basically, they trained the neural network to direct the drone to find a red camping chair, then move the chair to different environments, under different lighting conditions. , at different times of the year and at different distances to see if the drone can still find the chair. The authors wrote: “The main conceptual driver of our work is not the causality in the abstract; rather, it is task understanding, that is, assessing whether a neural model understands the task given off-line data that is not labeled with height.”
Daniela Rus, director of CSAIL and one of the co-authors, speak: “Our tests demonstrate that we can effectively teach a drone to locate an object in the forest in the summer, and then deploy the model in the winter, with very different surroundings or even in urban environments, with various tasks such as search and tracking. ”
Essentially, Dr. Hasani said, “they can generalize to situations they’ve never seen before.” Liquid neural networks can also “automatically capture the true cause and effect of the assigned task,” the authors write. This is “the key to the strong performance of liquid networks in distributional variations.”
The main advantage of liquid neural networks is their adaptability; neurons behave like worms (or neurons of other living things), reacting to real-world situations in real time. “They can change their basic equations based on the input they observe,” said Dr. Rus. speak Quantum journal.
Dr Rus further notes: “We are excited by the great potential of our robotic learning-based control approach, as it lays the groundwork to solve the problems that arise when mining. create in one environment and deploy in a completely different environment with no additional training… These flexible algorithms could one day support decision-making based on ever-changing data streams time, such as medical diagnostics and autonomous driving applications.”
Sriram Sankaranarayanana computer scientist at the University of Colorado, was impressed, speak Quantum journal: “The main contribution here is the stability and other nice properties brought into these systems by their sheer structure… They are complex enough to allow interesting things to happen, but not complicated. complex enough to lead to chaotic behavior.”
Alessio Lomuscioprofessor of AI safety in the Department of Computing at Imperial College London, was also impressed, tell MIT:
Strong learning and performance in unsecured tasks and scenarios are some of the key issues that machine learning systems and autonomous robots must conquer to move deeper into mission-critical applications. for society. In this context, the performance of liquid neural networks, a new brain-inspired model developed by authors at MIT, reported in this study is remarkable. If these results are confirmed in other experiments, the model developed here will contribute to making AI systems and robots more reliable, robust and efficient.
It’s easy to envision the plethora of drone applications where these could prove important, with autopilot another sensible use. But the MIT team is looking broader. “The results in this paper open up the possibility of certifying machine learning solutions for safety-critical systems,” said Dr. Rus. With all the discussion about the importance of ensuring that AI gives valid answers in healthcare use, as noted above, she specifically mentioned bringing medical diagnostic decision making as a decision for liquid neural networks.
“Everything that we do as a robotics and machine learning lab is [for] “Completely secure and safely and ethically deploy AI in our society, and we really want to stick to the mission and vision we have,” said Dr. Hasani. We should hope that other AI labs feel the same way.
Healthcare, like most parts of our economy, will increasingly use and even rely on AI. We will need AI that can not only give accurate answers but can also adapt to rapidly changing conditions, rather than pre-established data models. I don’t know if it will be based on a liquid neural network or something else, but we will want not only adaptability but also safety and ethics.
Last month I Written about Organoid Intelligence (OI), intends to approach AI using world structures more like our brains. Now the liquid neural network is based on the worm’s brain. It is exciting to me that after several decades of research, and perhaps for our silicon coatings, we are starting to move towards more biological approaches.
EQUAL Sayan Mitraa computer scientist at the University of Illinois, Urbana-Champaign, speak Quantum journal: “It’s pretty poetic in a way, suggesting that this study may be going full circle. Neural networks are evolving to the point where the very ideas we draw from nature may soon help us better understand nature.”
Kim is a former director of e-marketing at a Blues grand scheme, editor of The Late & Lamentations Tincture.ioand is now a regular THCB contributor.
