Intel's research laboratories reveal 31 research projects of artificial intelligence innovation, including three studies from Israel
Intel Labs, the group of Intel's research laboratories, revealed in recent days how the world of chips, electronics and AI is expected to look in the coming years. During the NeurIPS 2023 conference – the global event for developers, researchers and academic professionals focused on AI and computer vision technologies.
Intel Labs presented at the conference 31 papers and studies that deal with new models, methods and tools for artificial intelligence use cases intended for application in climate models, drug discovery and materials science. Two prominent studies of these are by Israeli researchers at Intel Laboratories.
A study in which the Israeli researcher took part Yaniv Gurvitz In collaboration with a research group at MILA presents ClimateSet, an innovative data set and tools for climate science research. The researchers have amassed a huge collection of climate data. In simple terms: it can be described as if you can predict the future of the Earth's climate in response to a future scenario of greenhouse gas emissions and aerosols. ClimateSet combines data from 36 different climate models from different groups of climate scientists around the world into one massive dataset.
The innovation in this research lies in the fact that instead of using data from only one climate model, as is usually done, ClimateSet brings together a wide variety of models that represent different predictions, which makes it possible to get a more complete, diverse and accurate picture of the future climate scenarios, while referring to the level of certainty in the prediction. Scientists use ClimateSet to explore a range of future scenarios, from the best case to the worst case.
The study also describes the use of "super-emulators" and advanced machine learning (ML) models, which will allow in the future to learn about the Earth's climate from a huge amount of data, which comes from different climate models.
"Super-Emulator" is a powerful software that uses ClimateSet to create fast and accurate predictions about the climate, and is orders of magnitude faster than traditional methods, which is very important for making timely and accurate decisions. For the purpose of comparison, today in order to simulate a climate model for the next decades (until the year 2100) for a single future scenario of greenhouse gas emissions and aerosols, many months (almost a year) of running on a supercomputer are required, which makes it impractical and not accessible. Accelerating the predictions will make them useful and accessible to the machine learning and climate science research communities.
"Predicting climate change for the next decades is a basic need for decision makers to adopt a correct climate policy. Intel and MILA carry out joint research of great importance for the prediction of climate changes, which are currently based on simulations of complex physical processes of climate models, and hence require a long running time that is impractical for the research groups. Through the development of causal (causal) machine learning methods, and deep learning-based emulators, we aim to speed up climate prediction by orders of magnitude, and make it accessible and useful. In addition, the research will enable climate prediction in response to greenhouse gas and aerosol emission scenarios, which so far have not been taken into account. The publication of the data and the open source, ClimateSet, the largest climate model database, is an important step towards fast and explainable climate prediction, and will also benefit the machine learning and climate science communities in dealing with other significant climate change tasks", Gurvitz points out.
A study written by Israeli researchers Estel Apello and Gabriela Ben Melech, is designed for XNUMXD virtual reality (VR) and aims to simplify XNUMXD panoramic image creation for artificial intelligence applications. The researchers developed a model that answers the challenge of creating realistic panoramic RGB images (red, green, blue) and their corresponding depth maps directly from textual instructions. The innovation lies in its ability to co-create detailed and accurate panoramic views in addition to the depth maps that are essential for immersive VR experiences.
A practical example: a VR application where you can create an environment from your imagination by text instructions, whether it is a relaxing natural environment or a home environment of the bedroom you dream of. All you have to do is enter a text message like "view of bedroom 360" and the model will generate a realistic panoramic image of the bedroom, with depth perception, allowing for a more immersive VR experience, allowing you to wander around the bedroom you created. You are welcome to try it yourself.
Another model the researchers developed specializes in improving the resolution of RGB images as well as depth maps. Say you have low resolution images of natural scenery for a VR game. The model can upgrade these images to high resolution, making the game environment more realistic and visually appealing.
The researchers note that these innovations push the boundaries of VR content creation and make it possible to create high-quality and realistic XNUMXD environments from simple text messages or low-resolution images. According to Intel, the models not only improve the VR user experience but also simplify and speed up the content creation process, which may be especially beneficial for VR developers and content creators in the entertainment, education and training sectors.
Discovering causes in large language models
Another study written by the Israeli researchers Raanan Yehezkel Rohker and Viniv Gurvitz, shows a connection between a central component of large language models and the causal structure hidden in the data. The researchers describe how this relationship can be used to discover the hidden causal structure, thus discovering direct and indirect cause and effect relationships between the variables processed by the models. By doing so, the researchers demonstrate how the causal structure explains the results of large language models, which will allow a better understanding of their way of working and improvement.
More of the topic in Hayadan:
