Advance in Neural Information Processing Systems
Training multilayer perceptrons with the extended Kalman algorithm
作者:
S Singhal，L Wu
关键词:
learning ； filtering ； estimation ； training ； optimization ； gradient descent ； Kalman filtering
摘要:
Recent degradation of coral reef habitats and the resulting loss in marine biodiversity has created an urgent need for improved characterization and monitoring including seafloor surveys. Most modern systems use sophisticated multibeam sonar and are prohibitively expensive to deploy. As part of a joint effort with The Nature Conservancy we have developed an inexpensive system for coral reef bathymetry estimation using inexpensive single-beam echosounders designed for fish finding. The depth tracking algorithms in these units are not well suited to coral reef habitats which often contain jagged structures and steep seafloor slopes. We propose to use a state-space model of the seafloor backscattering and track the depth with the extended Kalman filter. The initial results suggest that this approach produces better estimates of the depth, is intuitive to tune, has natural method of identifying and diagnosing problems, and provides a measure of confidence in the depth estimates.
在线下载