Personalized Active Learning

Most active learning (AL) methods rely on uncertainty sampling to gather new most informative information in an environment. As real-world environments can contain information unrelated to the tasks that the autonomous system needs to perform, uncertainty sampling can lead to useless information learned by the system through limited queries that could be used to learn task-related information. This can also negatively affect users' trust in the system. We develop AL methods that can choose queries that will generate the most important information related to the task that the user needs the autonomous system to perform. A major challenge in developing such AL methods is to determine if unknown data in the environment is related to the task without querying the user about the data. We are developing novel similarity metrics and leveraging LLMs to determine if an unknown data sample is related to the task. However, this led to another challenge, this objective of determining and choosing information most similar to the task, is opposite to uncertainty sampling to find the most dissimilar samples compared to the previously learned knowledge. We investigate hierarchical similarity metrics to find similar information on the task level, but dissimilar information on the lower sample-level knowledge.

Please contact Dr. Ali Ayub for more information

Relevant publications

Publications pdf

---