Computational Signal Acquisition

My research focus is on Computational Signal Acquisition. In contrast to classical signal acquisition approaches, my work relies heavily in computation. I believe that the future of sensing technology lies in incorporating computational methods throughout the design of acquisition systems. Sensing and acquisition systems aim to capture physical signals and reconstruct them digitally or extract useful information. Co-developing such systems with computational reconstruction and information extraction methods is paramount to achieving the potential of the sensing hardware.

The recently emerged area of Compressive Sensing (CS) has shown the way towards this paradigm. The breakthrough in CS was the realization that appropriate signal modeling and computational methods can significantly reduce the number of measurements required to acquire most signals we want to acquire, compared to what Nyquist-rate theorems dictated–hence the term “compressive.”  The success of this area has spawned a large body of work under the CS moniker, a significant fraction of which is computational but not really “compressive.”

The term “computational signal acquisition” aims to capture a broader set of goals and methods. A good analogy is computational imaging, a field using CS-like techniques even before the CS literature provided the theoretical foundation for such methods. The goal in computational imaging is not necessarily to reduce the number of measurements; in such systems measurements are inexpensive thanks to advances in CCD array technology.  Instead, the goals of computational imaging include post-capture focus adjustment, easier de-blurring, depth sensing and light-field capturing. My research aims to provide similar benefits to a wide variety of sensing systems. Fundamentally, there are four aspects one needs to investigate to systematically analyze a sensing system:

  • System Models: System models describe how the hardware operates and acquires the signal. They are critical in improving acquisition hardware and subsequent signal reconstruction. A system model captures aspects such as quantization, acquisition noise and non-linearities, and describes how the acquired data should be interpreted. Typically, but not always, the system model is introduced as an appropriate data fidelity criterion in the processing algorithm.
  • Signal Models: Signal models describe the kind of signals the hardware expects to acquire and describe how to resolve ambiguity in the data. For a model to be useful it should be computationally tractable, either exactly or approximately, and enforceable when processing the acquired data.
  • Information and Computation Scalability: The amount of information required by the application and the availability of computing power determines how much effort should be allocated in accurately sensing the signal. For example, we should expect that simple detection and estimation applications should require systems with lower sensing and processing complexity, compared to applications that require reconstruction of the whole signal.
  • Applications: While general principles are universal, it is critical to understand application-specific details. Real-life data test the accuracy of the models and the limits of the theory, significantly advancing our understanding. Co-developing theory and hardware for a particular application is the best way to identify how to redesign a sensing system to have the most impact.