An Active Learning Framework for Increasing Generalizability of Machine Learning Models

Recent years have witnessed a fast development of deep learning and significant success in practical application, such as object detection, recognition, image classification, and natural language processing. Deep learning has also been applied to space exploration, like the recognition of environmental features and classification of planets and supernova. The state-of-the-art deep learning methods require a huge amount of labeled data to train the neural networks, and the training process is computationally expensive and time-consuming. Although many learning models have achieved very impressive performance on benchmark datasets, they all face the issue of generalizability. Most machine learning models, even being well trained on very large datasets, suffer a big loss (4% to 10%) in accuracy for unseen data. Generalizability is still an open problem in machine learning since all learning models, once trained, have to be applied to new environments with new data that have not seen before.

This project aims to solve the generalization challenge of the learned models. We propose an intelligent learning framework by integrating the strategy of active and reinforcement learning to refine the model for unseen datasets. The overall idea is motivated by the observation of children learning to interact with their environment. The proposed framework is composed of three steps mimicking the learning process of children: Observe, practice, and improve, through which the machine can learn the knowledge incrementally from previous tasks and adapt it seamlessly in learning new tasks. Thus, the system can incrementally improve its performance during practical applications. This strategy has a potential impact on current leaning systems.

In addition to its scientific significance, the project has a great impact on NASA and Kansas economy. First, the project fulfills NASA’s strategic goal to advance technological innovations for NASA’s scientific missions. With large volumes of data collected from various missions of NASA, it is essential to develop new approaches for analysis and discovery by taking advantage of the fast development of artificial intelligence. Second, the study has a significant impact on the economic and scientific development of Kansas. It is reported that about one in five Kansas City jobs is predicted to have high exposure to AI, especially in computers, business, and finance. AI also has a significant impact on the manufacturing and agriculture of Kansas.

The project will also increase the level of education and research of Kansas. Many students want to take AI-related courses, and the project will help the involved students obtain hands-on research experience in AI. Many professors and researchers in Kansas are starting to use machine learning in their research. The proposed solution will benefit all researchers working in machine learning and enable many other related research topics. Moreover, the project will collaborate with NASA JPL and a local high-tech company. These collaborations will significantly increase Kansas’s existing research strength in artificial intelligence, as well as our competitiveness in obtaining external research funding.

This is a mature technology (TRL 7+) — the realistic path in is usually NASA's Technology Transfer Program: licensing an existing NASA patent, or a Space Act Agreement to use NASA facilities/expertise directly. NASA also runs a startup licensing program with no upfront fee for companies formed to commercialize a specific NASA technology.

None of these are guaranteed paths for this specific project — TechPort itself doesn't have an "apply" button. Reaching out to the contact(s) above with a specific question is usually the fastest way to find out what's actually open.