Learning to Communicate with Intent: An Introduction
This addresses the inefficiency of classical communication systems for goal-oriented tasks, offering a novel approach that could enhance applications like image transmission and gaming, though it appears incremental in adapting existing learning methods to communication.
The paper tackles the problem of designing communication systems that transmit messages based on the end-goal rather than exact reconstruction, proposing a framework for joint learning of communication and tasks like classification and reinforcement learning. Results show significant improvements over traditional joint source-channel coding, with close-to-upper-bound performance in reinforcement learning even at low SNRs.
We propose a novel framework to learn how to communicate with intent, i.e., to transmit messages over a wireless communication channel based on the end-goal of the communication. This stays in stark contrast to classical communication systems where the objective is to reproduce at the receiver side either exactly or approximately the message sent by the transmitter, regardless of the end-goal. Our procedure is general enough that can be adapted to any type of goal or task, so long as the said task is a (almost-everywhere) differentiable function over which gradients can be propagated. We focus on supervised learning and reinforcement learning (RL) tasks, and propose algorithms to learn the communication system and the task jointly in an end-to-end manner. We then delve deeper into the transmission of images and propose two systems, one for the classification of images and a second one to play an Atari game based on RL. The performance is compared with a joint source and channel coding (JSCC) communication system designed to minimize the reconstruction error of messages at the receiver side, and results show overall great improvement. Further, for the RL task, we show that while a JSCC strategy is not better than a random action selection strategy even at high SNRs, with our approach we get close to the upper bound even for low SNRs.