Deep Error Prediction

Areté Associates proposes to use recent advances in deep learning to aid operators and analysts currently processing multiple streams of data in real-time, significantly reducing the cognitive burden of human-robot teaming. The “black-box� nature of Automatic Target and Recognition (ATR) algorithms gives rise to this burden and without meaningful error bounds on these ATR outputs the human operator cannot build trust in the algorithms. The Areté Learned Likelihood Statistical Error Estimation Network (ALLSEEN) applies advances in deep artificial-neural-network methods to estimate the confidence interval of a detection. ALLSEEN is able to simultaneously detect, localize and characterize objects in a scene while outputting the covariance matrix of these inferences. During the course of Phase I, Areté will adapt ALLSEEN to operate on an UAV dataset and demonstrate the feasibility of the technique. During Phase II, Areté will further develop and demonstrate a real-time prototype of ALLSEEN to enhance the ISR mission with validated mission utility on an unmanned system.