Principal Investigator: Assistant Professor Wang Xinchao, ECE

Endeavor to develop a robust, dependable, and generalizable framework to tackle incremental learning in the wild, and to introduce generic solutions to handle catastrophic forgetting, thereby enabling incremental learning to be readily deployable to a large domain of scenarios. We aim to accomplish the following sub-objectives:
1. Establishing a principled framework to enable general-purpose incremental learning, which is not limited to Euclidean data like images, but applicable to non-Euclidean data like point clouds as well.
This will involve exploring self-supervision to further enhance incremental learning performance.
2. Equipping incremental learning approaches with the capability to handle multi-modality data as input, which is a common setup in real-world problems but unfortunately barely investigated by existing incremental learning schemes.
3. Devising compact and multi-task incremental learning approaches that allow for tackling multiple tasks jointly under a lightweight incremental learning architecture.

The completion of these three sub-objectives will truly enable a general-purpose incremental learning that is free of application- and setup-wise constraints, and hence significantly boosts the robustness, reliability, and performance of incremental learning models. We will test the proposed incremental learning approaches on standard benchmarks, and also collect a large-scale dataset with in-the-wild data to showcase the competitiveness of our approaches against the state of the art.