Quantization in acquisition and computation networks

Abstract

In modern systems, it is often desirable to extract relevant information from large amounts of data collected at different spatial locations. Applications include sensor networks, wearable health-monitoring devices and a variety of other systems for inference. Several existing source coding techniques, such as Slepian-Wolf and Wyner-Ziv coding, achieve asymptotic compression optimality in distributed systems. However, these techniques are rarely used in sensor networks because of decoding complexity and prohibitively long code length. Moreover, the fundamental limits that arise from existing techniques are intractable to describe for a complicated network topology or when the objective of the system is to perform some computation on the data rather than to reproduce the data. This thesis bridges the technological gap between the needs of real-world systems and the optimistic bounds derived from asymptotic analysis. Specifically, we characterize fundamental trade-offs when the desired computation is incorporated into the compression design and the code length is one. To obtain both performance guarantees and achievable schemes, we use high-resolution quantization theory, which is complementary to the Shannon-theoretic analyses previously used to study distributed systems. We account for varied network topologies, such as those where sensors are allowed to collaborate or the communication links are heterogeneous. In these settings, a small amount of intersensor communication can provide a significant improvement in compression performance. As a result, this work suggests new compression principles and network design for modern distributed systems. Although the ideas in the thesis are motivated by current and future sensor network implementations, the framework applies to a wide range of signal processing questions. We draw connections between the fidelity criteria studied in the thesis and distortion measures used in perceptual coding. As a consequence, we determine the optimal quantizer for expected relative error (ERE), a measure that is widely useful but is often neglected in the source coding community. We further demonstrate that applying the ERE criterion to psychophysical models can explain the Weber-Fechner law, a longstanding hypothesis of how humans perceive the external world. Our results are consistent with the hypothesis that human perception is Bayesian optimal for information acquisition conditioned on limited cognitive resources, thereby supporting the notion that the brain is efficient at acquisition and adaptation.