A Multistage Architecture for Statistical Inference with Stochastic Signal Acquisition

We describe a statistical inference approach for designing signal acquisition interfaces and inference systems with stochastic devices. A signal is observed by an array of binary comparison sensors, such as highly scaled comparators in an analog-to-digital converter, that exhibit random offsets in their reference levels due to process variations or other uncertainties. These offsets can limit the performance of conventional measurement devices. In our approach, we build redundancy into the sensor array and use statistical estimation techniques to account for uncertainty in the observations and produce a more reliable estimate of the acquired signal. We develop an observational model and find a Cramér-Rao lower bound on the achievable square error performance of such a system. We then propose a two-stage inference architecture that uses a coarse estimate to select a subset of the sensor outputs for further processing, reducing the overall complexity of the system while achieving near-optimal performance. The performance of the architecture is demonstrated using a simulated prototype for parameter estimation and symbol detection applications. The results suggest the feasibility of using unreliable components to build reliable signal acquisition and inference systems.