RAPID HIGH-DIMENSIONAL SEMANTIC SEGMENTATION WITH ECHO STATE NETWORKS

Recurrent Neural Networks have largely been explored for low-dimensional time-series tasks due to their fading memory properties, which is not needed for feed-forward methods like the Convolutional Neural Network. However, benefits of using a recurrent-based neural network (i.e. reservoir computing) for time-independent inputs includes faster training times, lower training requirements, and reduced computational burdens, along with competitive performances to standard machine learning methods. This is especially important for high-dimensional signals like complex images. In this report, a modified Echo State Network (ESN) is introduced and evaluated for its ability to perform semantic segmentation. The parallel ESN containing 16 parallel reservoirs has an image processing time of 2 seconds with an 88% classification rate of 3 classes, with no prior feature extraction or normalization, and a training time of under 2 minutes.

Citation: S. Gardner, M. R. Haider, J. Smereka, P. Jayakumar, K. Kulkarni, D. Gorsich, L. Moradi, and V. Vantsevich, “Rapid High-dimensional Semantic Segmentation With Echo State Networks”, In Proceedings of the Ground Vehicle Systems Engineering and Technology Symposium (GVSETS), NDIA, Novi, MI, Aug. 10-12, 2021.