A Markovian Error Model for False Negatives in DNN-based Perceptron-Driven Control Systems
Kruttidipta Samal, Thomas Walton, Hoang-Dung Tran, Marilyn Wolf. 2022. A Markovian Error Model for False Negatives in DNN-based Perceptron-Driven Control Systems. Accepted, International Neural Network Society Workshop on Deep Learning Innovations and Applications (INNS DLIA 2023). IEEE.
Abstract
This paper presents an improved Markovian error model for Deep Neural Network (DNN) based perception in autonomous vehicles and other perception-driven control systems. Many modern autonomous systems rely on DNN-driven perception-based control/planning methodologies such as autonomous navigation, where the perception errors significantly affect the control/planning performance and the systems’ safety. The traditional independent, identically-distributed (IID) perception error model is inadequate for perception-based control/planning applications because image sequences supplied to a DNN-based perception module are not independent in the real world. Based on this observation, we develop a novel Markov model to describe the error behavior of a DNN perception model—an error in one frame is likely to signal errors in successive frames, effectively reducing sample rate for the control command. We evaluate the effect of Markovian perception errors on a drone-control simulator and show that the Markovian error model provides a better estimate of control performance than does a traditional independent, identically-distributed (IID) model.
Under the supervision of Dr. Mohammad R. Hasan, I investigated the impact of algorithmic bias in state of the art CNNs. Using ScoreCAM, a method of interpreting the gradients of a neural network in a meaningful way, we were able to address this bias. Our findings proved that deep beliefs held by networks like MobileNetV3 and ResNet-50 are fundamentally flawed. By artificially injecting confusing objects into images of flowers, weaknesses in these networks became apparent. While Deep Learning is a fast progressing field, there are still many questions to be answered about the inner workings of these networks. This research exposes a weakness in the current paradigm of the field, opening up new opportunities to make networks even better.