Deep lifelong learning with privacy preserving memory

Description of field of research:

As a cornerstone of modern artificial intelligence (AI), deep learning (DL) has achieved state-of-the-art performances in various applications, such as visual understanding for autonomous driving, speech recognition, and medical diagnosing. Most methods are restricted to the traditional view of a static learning process on fixed datasets.

Lifelong learning (LL) or continual learning (CL) aims to incrementally learn and update the deep learning models on streaming data, making AI evolve in a lifelong circle even after deployment, like humans. The model will continually learn new knowledge from the new datasets/tasks in the stream while maintaining the performance of all tasks. To avoid forgetting in the long-term dynamic learning process, LL or CL methods can maintain a memory to store the representative data samples seen during training.

With the general application of AI and DL in people's daily life, the privacy issue becomes a big challenge for responsible AI and DL. In this project, we will investigate how to build a privacy-preserving memory for lifelong learning from data anonymization and compression perspectives. The developed method will maintain the memory buffer containing data, in which the privacy-sensitive information is suppressed, but the essential semantic information for DL (in LL/CL) can be preserved.