Intel showcased its laboratory's research results in the field of artificial intelligence technology at the NeurIPS 2023 conference and used this to promoteThe vision of ubiquitous artificial intelligence.
During the conference, Intel published a total of 31 research papers, including 12 main conference papers, 19 workshop papers and on-site presentations. The research focused on new models, methods and tools for the application of artificial intelligence in scientific fields, as well as graph learning and multi-modal generative AI. It also included artificial intelligence algorithms and optimization techniques for AI use cases such as climate modeling, drug development and materials science.
In addition, Intel Labs also held the Artificial Intelligence Accelerated Materials Exploration (AI4Mat) workshop, providing a platform for artificial intelligence researchers and materials scientists to address the challenges of artificial intelligence-driven materials exploration and development.
The research papers published this time are as follows:
Scientific Artificial Intelligence:
• Brain encoding models:The model, developed with researchers from the University of Texas at Austin, helps predict brain responses and provides insights into the brain's multimodal processing capabilities.
• ClimateSet:Developed in collaboration with the Quebec Institute for Artificial Intelligence (Mila), this large-scale climate model dataset for machine learning can rapidly predict new climate change scenarios and lay the foundation for the machine learning (ML) community to build disruptive and innovative climate-centric applications.
• HoneyBee:State-of-the-art large-scale language models developed with Mila to help researchers gain faster understanding of materials science.
Multimodal Generative Artificial Intelligence:
• COCO-Counterfactuals:A multimodal technique for generating synthetic counterfactual data can reduce incorrect statistical biases in pre-trained multimodal models, helping to improve the performance of AI models in many downstream tasks (such as image text retrieval and image recognition).
• LDM3D-VR:A potential diffusion model for 3D virtual reality (VR) simplifies the generation of 3D video using artificial intelligence applications.
• CorresNeRF：Image rendering method for reconstructing 2D scenes from 3D images using neural radiance fields.
Improving AI performance:
• DiffPack:A generative artificial intelligence approach to protein modeling that helps ensure that the generated 3D structures reflect the protein's true structural properties.
• InstaTune:A method for generating a super network during the fine-tuning phase to reduce the overall time and computing resources required for network-attached storage (NAS).
Graph Learning:
• A*Net:The industry's first path-based approach to knowledge graph reasoning on millions of datasets can be extended to datasets beyond computational capacity and improve the accuracy of large language models (LLMs).
• ULTRA:The industry's first foundational model for knowledge graph reasoning, and a new approach to learning general and transferable graphs and their relationships.
• Perfograph:A novel program representation based on compiler graphs can capture numerical information and composite data structures to improve the ability of ML methods to reason about programming languages.
