We created a new test for artificial intelligence driving systems to see if they truly understand how to follow traffic rules. The results show that current AI systems are good at some basic tasks, but struggle with more complex situations, and can become better drivers by learning from our new test.
Read Full Paper →Scientists studied a problem with computer vision models, which can't always accurately identify key events in images. By changing how these models process data, they developed a new approach that significantly improved performance and could have practical uses like better image recognition for self-driving cars or security systems.
Read Paper →We created a new video analysis tool called VoCap that can identify objects in videos and provide detailed descriptions of what's happening in the scene, using information from text prompts or labels. This technology has the potential to improve how we understand and work with videos in various fields, such as video editing, content creation, and research.
Read Paper →Scientists tested how well computer models can predict patient outcomes using data from different sources, like medical scans and gene tests, when the model has only been trained on one type of cancer. They found that these models don't work as well in predicting outcomes for other types of cancer, and proposed new ways to make them more accurate by improving how they use weak signals and combining different types of data.
Read Paper →The researchers developed a new system that helps identify unusual patterns in images by using three separate "memory banks" that store different types of information. This system, called TMUAD, can improve the detection of anomalies and has already shown to be highly effective in several real-world applications, such as inspecting industrial equipment or medical imaging.
Read Paper →
