RI: Small: SM-An Active Approach for Data Engineering to Improve Vision-Language Tasks

Arizona State University, Scottsdale AZ

Investigators

Abstract

Intelligent systems that can robustly process vision and language data are necessary to enable integrated AI applications (such as automated driving, robotic home assistant, etc.) and improve quality of life. However, such systems typically operate in open and highly uncertain environments for which physical and geometric understanding, semantic robustness, and conducting hypothetical reasoning become essential. This project will result in a publicly available software suite that can assist with training and validating robust Vision and Language (V&L) systems. In particular, the resulting semantic transformations will be packaged as an API service that companies and universities could quickly utilize. The resulting benchmark challenges will be made publicly available for further V&L research. Finally, the proposed study will stimulate educational activities at ASU in training graduate and undergraduate students in AI/ML/CV/NLP with a "post-dataset era'" vision. The project will also train 2 Ph.D. students and several master-with-thesis students, develop a new seminar course, recruit underrepresented minority participants at all levels, and reach K-12 students with modules that explain the challenges in developing robust intelligent systems. Robust intelligent systems such as home assistant robots fundamentally depend on highly correlated vision and language systems and fine-grained data alignment. Even though the existing approaches demonstrate success on carefully collected benchmarks, it is not sufficient to establish robustness, reliability, and out-of-distribution generalization for them to be deployed in real-world applications. The project will conduct a systematic study on intelligent and active data engineering to boost their performance and robustness. By investigating a novel and active perspective towards vision and language data engineering, the project will address the following three fundamental research tasks: 1) development of data generators to hallucinate training data from existing ones with low-level vision; 2) with hypothetical actions, and 3) design of training paradigms incorporating the new data generated with the goal of increasing the ultimate systems' generalization capability and robustness. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.