MIT Lincoln Laboratory · 2012 to 2015
Radar & Sensor Signal Processing
Detection and estimation against noisy, nonstationary sensor data, and the origin of everything I think about signal.
Role · Technical Staff.
The problem
Real instrument data does not behave like the data in a textbook. It drifts, the noise is not stationary, and a great deal of what looks like structure is an artifact of the measurement rather than a property of the world. On top of that the loss function is asymmetric: a false alarm and a missed detection are not equally bad, and pretending otherwise produces a system nobody can use.
Approach
- 01
Applied spectral and time-domain methods to sensor data, working the detection and estimation problem end to end rather than at the level of an isolated algorithm.
- 02
Treated the measurement chain as part of the model, because the fastest way to a wrong answer is to trust the instrument.
- 03
Set thresholds against the actual cost of each error type instead of against a symmetric accuracy metric.
Outcome
Three years of it, alongside the BU master's. The habit it left behind is the useful part: assume apparent structure is noise until the data forces you to say otherwise.