Evaluating Fisher Information From Data for Task-Driven Data Compression
Mark L. Fowler, Mo Chen · 2006
Sensor networks and other multi-sensor systems collect data upon which estimations are based. How a particular sensor's data is to be used within the sensor system depends on the quality of that sensor's data relative to the other sensor's data. Because Fisher information (FI) is a natural way to assess the quality of a sensor's data, it is desirable to be able to numerically compute the FI for a collected set of data for a specific estimation task. We explore this issue for the specific scenario of FI-driven data compression for a sensor system tasked with estimating an RF emitter's location. The data compression scheme uses sub-band coding and therefore for FI-driven data compression it is important to numerically assess the FI of each filter bank output sample. However, as we demonstrate, an orthogonal filter bank well-suited to the compression task seems ill-suited to the FI-assessment task. Alternatively we demonstrate that the STFT is well-suited to FI-assessment but - as is well known - is ill-suited to the compression task. This leads to a hybrid structure for the sub-band coding scheme that uses the STFT for time-frequency domain assessment of the FI but retains the orthogonal filter bank for the compression processing.