The need for regularization in feedforward systems

Post date: Jun 04, 2017 4:23:28 PM

A feedforward system cannot be correctly self-evaluated and hence self-corrected. Herein, a system is defined as the posterior probabilistic mapping , where is the input/observation and the output is the expected latent state

. Hence, it is tempting to evaluate the system performance as the variance of the posterior. However, if the true (unknown) input-output relationship

deviates from the system, then such an evaluation is inaccurate. For instance, IMU-based systems often experience location drifts overtime, and no amount of filtering can recover the true location.

One way to handle the above problem is to expand the input space and hope that the system model

 matches the real-world one . For examples, augmenting the IMU data with video data. Alternatively, the system model can be regularized to match that of the real-world model. The periodic assumption on motions in [3] helps reducing the location drift. The Huber uncertainty model in [1,2] is a form of regularization.

[1] Long N. Le, Douglas L. Jones. "Guided-Processing Outperforms Duty-Cycling for Energy-Efficient Systems.'' IEEE Transactions on Circuits and Systems I, 2017.

[2] Long N. Le, Douglas L. Jones. "Feature-Sharing in Cascade Detection Systems with Multiple Applications." IEEE Journal of Selected Topics in Signal Processing, 2017.

[3] Latt, W. T., Veluvolu, K. C., & Ang, W. T. (2011). Drift-free position estimation of periodic or quasi-periodic motion using inertial sensors. Sensors, 11(6), 5931-5951.