Kagan Tumer's Publications

Display Publications by [Year] [Type] [Topic]


Sequence Recognition by Input Anticipation. K. Tumer and J. Ghosh. In Proceedings of the Seventh International Conference on Industrial and Engineering Applications of Artificial Intelligence and Expert Systems, Austin, TX, June 1994.

Abstract

Processing time-dependent information is crucial in a multitude of classification tasks. This paper introduces a new neural network called the anticipation network, which fuses two simple recurrent networks for more efficient feature extraction and classification of spatiotemporal signals. The network has two sets of outputs: the first set classifies the patterns, while the second one trains "anticipation" units whose goal is to predict future inputs. The correlation between these outputs adds constraints to shared hidden nodes, thereby resulting in better feature selection. The ability to anticipate the input sequence is thus used to either classify, generate or complete sequences. Simulation results are provided for a simple frequency detection task, classification of artificial (multidimensional) sonograms, and the detection and classification of real underwater acoustic signals. The results show that anticipation networks not only classify patterns more accurately, but are able to make a correct decision at an earlier time as compared to other recurrent networks.

Download

(unavailable)

BibTeX Entry

@inproceedings{tumer-ghosh_ieaai94,
       author={K. Tumer and J. Ghosh},
       title="Sequence Recognition by Input Anticipation",
       booktitle="Proceedings of the Seventh International Conference on 
	Industrial and Engineering Applications of Artificial Intelligence
	and Expert Systems",
       month={June},
	address={Austin, TX},
	abstract={Processing time-dependent information is crucial in a multitude of classification tasks. This paper introduces a new neural network called the anticipation network, which fuses two simple recurrent networks for more efficient feature extraction and classification of spatiotemporal signals. The network has two sets of outputs: the first set classifies the patterns, while the second one trains "anticipation" units whose goal is to predict future inputs. The correlation between these outputs adds constraints to shared hidden nodes, thereby resulting in better feature selection. The ability to anticipate the input sequence is thus used to either classify, generate or complete sequences. Simulation results are provided for a simple frequency detection task, classification of artificial (multidimensional) sonograms, and the detection and classification of real underwater acoustic signals. The results show that anticipation networks not only classify patterns more accurately, but are able to make a correct decision at an earlier time as compared to other recurrent networks.},
	bib2html_pubtype = {Other Conference Papers},
	bib2html_rescat = {Other Topics},
       year={1994}
     }

Generated by bib2html.pl (written by Patrick Riley ) on Tue Jun 26, 2018 19:10:42