In goal-oriented communications, indirect source coding involves two sources: a semantic source $X$ and an observation $Z$, which is the only source available at the encoder. In this paper, we consider this framework for classification, where the decoder only needs to reconstruct $X$. We argue that the gap between the indirect rate-distortion function and the performance of practical learned coding schemes, such as variational information bottleneck (VIB) and variational autoencoder (VAE) architectures, comes from three effects: the latent representation, the choice of the deep neural network (DNN) architecture and training loss, and whether the encoder is aware of the task. To quantify these effects, we introduce four operational rate-distortion functions that account for the DNN architecture and the training loss, and we establish a hierarchy among them. We then derive these functions for a Gaussian mixture source, where $X$ is the mixture state. Numerical results, both for the Gaussian mixture and for practical coding schemes on the CIFAR-100 dataset, quantify the rate loss incurred by each effect.

翻译：暂无翻译