Modern deep learning, while effective in tackling clean and curated datasets, is often very brittle to domain and distribution shifts [76].

Artificial intelligence might now be solving advanced math, performing complex reasoning, and even using personal computers, but today's algorithms could still learn a thing or two from microscopic worms. Liquid AI, a startup spun out of MIT, will today reveal several new AI models based on a novel type of "liquid" neural network that has the potential to be more efficient, less power-hungry, and more transparent than the ones that underpin everything from chatbots to image generators to facial recognition systems. Liquid AI's new models include one for detecting fraud in financial transactions, another for controlling self-driving cars, and a third for analyzing genetic data. The company touted the new models, which it is licensing to outside companies, at an event held at MIT today. The company has received funding from investors that include Samsung and Shopify, both of which are also testing its technology.

There is an ever-growing zoo of modern neural network models that can efficiently learn end-to-end control from visual observations. These advanced deep models, ranging from convolutional to patch-based networks, have been extensively tested on offline image classification and regression tasks. In this paper, we study these vision architectures with respect to the open-loop to closed-loop causality gap, i.e., offline training followed by an online closed-loop deployment. This causality gap typically emerges in robotics applications such as autonomous driving, where a network is trained to imitate the control commands of a human. In this setting, two situations arise: 1) Closed-loop testing in-distribution, where the test environment shares properties with those of offline training data. 2) Closed-loop testing under distribution shifts and out-of-distribution. Contrary to recently reported results, we show that under proper training guidelines, all vision models perform indistinguishably well on in-distribution deployment, resolving the causality gap. In situation 2, We observe that the causality gap disrupts performance regardless of the choice of the model architecture. Our results imply that the causality gap can be solved in situation one with our proposed training guideline with any modern network architecture, whereas achieving out-of-distribution generalization (situation two) requires further investigations, for instance, on data diversity rather than the model architecture.

MIT researchers have developed a type of neural network that learns on the job, not just during its training phase. These flexible algorithms, dubbed "liquid" networks, change their underlying equations to continuously adapt to new data inputs. The advance could aid decision making based on data streams that change over time, including those involved in medical diagnosis and autonomous driving. "This is a way forward for the future of robot control, natural language processing, video processing--any form of time series data processing," says Ramin Hasani, the study's lead author. "The potential is really significant."

In the animal kingdom, brains come in all shapes and sizes. So, in a new machine learning approach, engineers did away with the human brain and all its beautiful complexity--turning instead to the brain of a lowly worm for inspiration. Turns out, simplicity has its benefits. The resulting neural network is efficient, transparent, and here's the kicker: It's a lifelong learner. Whereas most machine learning algorithms can't hone their skills beyond an initial training period, the researchers say the new approach, called a liquid neural network, has a kind of built-in "neuroplasticity." That is, as it goes about its work--say, in the future, maybe driving a car or directing a robot--it can learn from experience and adjust its connections on the fly.

MIT researchers have developed a type of neural network that learns on the job, not just during its training phase. These flexible algorithms, dubbed "liquid" networks, change their underlying equations to continuously adapt to new data inputs. The advance could aid decision making based on data streams that change over time, including those involved in medical diagnosis and autonomous driving. "This is a way forward for the future of robot control, natural language processing, video processing -- any form of time series data processing," says Ramin Hasani, the study's lead author. "The potential is really significant."

"This is a way forward for the future of robot control, natural language processing, video processing -- any form of time series data processing," says Ramin Hasani, the study's lead author. "The potential is really significant." The research will be presented at February's AAAI Conference on Artificial Intelligence. In addition to Hasani, a postdoc in the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), MIT co-authors include Daniela Rus, CSAIL director and the Andrew and Erna Viterbi Professor of Electrical Engineering and Computer Science, and PhD student Alexander Amini. Other co-authors include Mathias Lechner of the Institute of Science and Technology Austria and Radu Grosu of the Vienna University of Technology.