Meta-learning to transfer experience from learned tasks to new tasks in multi-task semantic communication for autonomous vehicles

Investigate meta-learning techniques to enable sharing of experience from tasks learned within the multi-task oriented semantic communication framework for connected and autonomous vehicles to new tasks, so that the framework can adapt knowledge from image reconstruction and classification of road traffic signs transmitted over vehicle–satellite–vehicle links to previously unseen tasks.

Background

The paper proposes a multi-task semantic communication framework for connected and autonomous vehicles in which a convolutional autoencoder encodes road traffic sign images and task-oriented decoders perform image reconstruction and classification after transmission via a vehicle–satellite–vehicle link. The approach demonstrates robustness in low SNR regimes and significant bandwidth savings compared to conventional QAM-16 baselines.

In the conclusion, the authors point to a gap regarding how knowledge acquired from the currently addressed tasks can be leveraged for new tasks. They explicitly state that sharing the experience from learned tasks to new tasks via meta-learning remains to be addressed, identifying a clear open direction for extending the framework's generalization across tasks.