Vector Researchers Present AI-Powered Search Engine for Data at CVPR 2020
Researchers from the Vector Institute for Artificial Intelligence and the University of Toronto presented Neural Data Server (NDS), a novel AI-powered search engine, at the CVPR 2020 conference. Developed by Sanja Fidler, David Acuna, and Xi Yan, NDS addresses the significant computational and time costs associated with pre-training deep neural networks. Unlike traditional search engines that simply index datasets, NDS uses machine learning to identify and recommend the most relevant subsets of publicly available data for specific user models. This process involves running lightweight local models to generate statistics, which are then analyzed by NDS to filter out irrelevant information, such as excluding vehicle images for fashion applications. By reducing the volume of data required for training, NDS significantly lowers resource demands, making high-performance model development more accessible to startups and academic researchers with limited infrastructure. The tool protects user privacy by not storing actual data, only processing statistical metadata. This innovation aims to democratize access to advanced AI capabilities by streamlining the critical pre-training phase, thereby saving money and compute power while maintaining model performance.
Wire timeline
Vector Researchers Present AI-Powered Search Engine for Data at CVPR 2020
Researchers from the Vector Institute for Artificial Intelligence and the University of Toronto presented Neural Data Server (NDS), a novel AI-powered search engine, at the CVPR 2020 conference. Developed by Sanja Fidler, David Acuna, and Xi Yan, NDS addresses the significant computational and time costs associated with pre-training deep neural networks. Unlike traditional search engines that simply index datasets, NDS uses machine learning to identify and recommend the most relevant subsets of publicly available data for specific user models. This process involves running lightweight local models to generate statistics, which are then analyzed by NDS to filter out irrelevant information, such as excluding vehicle images for fashion applications. By reducing the volume of data required for training, NDS significantly lowers resource demands, making high-performance model development more accessible to startups and academic researchers with limited infrastructure. The tool protects user privacy by not storing actual data, only processing statistical metadata. This innovation aims to democratize access to advanced AI capabilities by streamlining the critical pre-training phase, thereby saving money and compute power while maintaining model performance.
Vector Institute for Artificial Intelligence