A deep neural network that currently treats data transmission as an opaque black box needs to be cracked open so engineers can see exactly how it works. Modern communication systems—from mobile phones to satellite links—rely on sending compressed information across noisy channels. Recent machine-learning designs have shown impressive performance by combining compression and transmission into a single neural network, but these systems are so inscrutable that engineers cannot easily adapt them, make them computationally efficient, or extend them to more complex settings like two-way conversations or multi-user networks. This project will systematically compare two ways of designing these systems: one rooted in information theory (the mathematical limits of communication) and one rooted in machine learning. By studying a sequence of increasingly complex signal types and channel conditions, the researchers aim to reveal why neural-network-based designs sometimes outperform traditional methods and where they fall short. If successful, the work could lead to communication systems that are more robust, less demanding on memory and processing power, and easier to deploy in real-world scenarios such as emergency response networks or satellite constellations. This is fundamental science—it will not produce a market-ready product, but it addresses a core bottleneck that currently prevents a promising class of communication technologies from leaving the lab.
View original technical description
Recent successes in the design of joint source-channel coding (JSCC) schemes based on machine learning techniques have revitalized the research interest in JSCC. Despite initial successes, there are many critical open issues that prevent their adoption in wider practice. The most significant issue is that the end-to-end designs based on deep neural networks (DNNs) hide the underlying operations, and in turn, they provide little insight, and are less interpretable. Such considerations are particularly important in communication systems for several reasons, including computation efficiency, flexibility in applying the schemes, and easiness of extending to channels with feedback and multi-user settings. Given the issues mentioned above, the objective of the proposed research is to significantly improve our understanding of the underlying mechanism of ML-based JSCC approaches, such that they become computationally efficient, less storage-hungry, more adaptive, more robust, and easily generalizable to complex communication settings. The key challenge here is that DNNs are known to be difficult to interpret since they combine multiple functions via end-to-end training. We propose to conduct comparative studies of the JSCC designs from two related but distinctive perspectives: an information-theoretic perspective and a machine-learning perspective. By studying a sequence of more and more complex vector Gaussian sources and channel scenarios, and by contrasting and comparing the schemes designed using these two different perspectives, we aim to develop new insights and simplification of the ML-based JSCC coding scheme designs. The three thrusts of the proposed research are: 1) Study the signal representation and neural network interface in point-to-point communication settings, and understand the sources of the performance gain and graceful degradation; 2) Study neural JSCC in feedback channels and multi-user communication settings; 3) Study generative models in neural JSCC, particularly the Gaussian diffusion models.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know