FEATURED COMPUTER VISION

AI Networks Can Now Undo Distortions and Restore Data

From arXiv • Latest Research

A close up of a control panel with buttons and knobs

Photo by Eric Prouzet on Unsplash • Unsplash

Scientists developed a new way to solve problems with data that has been altered or distorted in some way, by creating a type of artificial neural network that can undo these changes. This breakthrough could have significant benefits for tasks like image and video restoration, as well as applications where precise control over generative outputs is needed without requiring extensive retraining.

Read Full Paper →

AI Learns to Forget Less and Adapt Faster

Scientists developed a new way to update pre-trained models so they can learn from multiple tasks without forgetting what they already know. This approach, called Share, could help make artificial intelligence systems more efficient and effective by allowing them to adapt to new situations without needing to be completely retrained.

Read Paper →

AI Models Get Better at Reading Your Mind With Help From Multiple Camera Views

We created a new system that helps AI models understand 3D scenes by combining information from multiple images taken from different viewpoints, allowing them to reason about objects and their relationships more accurately. This innovation has the potential to make it easier for people to use AI in real-world applications, such as virtual reality or robotics, where understanding complex scenes is crucial.

Read Paper →

a bird flying through a blue sky on a sunny day

Photo by Aleksandrina Andreeva on Unsplash • Unsplash

SwimBird Switches Gears for Better AI Understanding

Scientists developed a new type of artificial intelligence system called SwimBird, which can switch between different ways of thinking - using text only, looking at images only, or combining both. This allows SwimBird to be more flexible and better at tasks that require understanding and reasoning about complex visual information, while still being good at traditional text-based logic tasks.

Read Paper →

Robots Get Smarter at Working Together With AI

Researchers created a system to help multiple robots work together to complete tasks that involve understanding scenes, answering questions, and manipulating objects in real-world settings. This system improves how well robots can communicate with each other and gather information they need to succeed at these tasks, making it more efficient and effective for them to work together.

Read Paper →

AI Learns to "See" Spaces Better Than Ever Before

The researchers created a new system called GeoThinker that helps large language models better understand and use visual information from 3D spaces to make decisions. This improved ability to integrate spatial structures is important because it enables machines to more accurately perceive and navigate complex environments, such as streets for self-driving cars or environments for tasks like embodied referring.

Read Paper →

AI Learns to Read Human Intentions From Data

We created a system called InterPrior that helps robots learn to interact with people in different situations by imitating human behavior and adapting to new tasks. This system is important because it enables robots to learn how to handle a wide range of objects and interactions, which could lead to more useful and practical applications in real-world settings.

Read Paper →

Artificial Intelligence Learns to Look Closely Before Making Up Its Mind

Scientists created a new system called V-Retrver that helps large language models look at images more carefully before making decisions about what they've found. This system improves the accuracy of searches and allows the model to focus on specific parts of an image that might be relevant,.

Read Paper →