Feature Extraction and Clustering of Tactile Impressions with Connectionist Models
Abstract
We study the feasibility of adaptive pattern recognition of robotic tactile impressions using connectionist models. This paper presents interim simulation results of coupled back-error propagation (BEP) networks that (i) extract relative gradient features via data compression, (ii) clusters families of grey-scale patterns constrained by geometry, size, and activation levels, and (iii) classifies these surface profiles to pre-specified categories. The constraints imposed on the training data are designed to capture the essence of tactile patterns and force the artificial neural systems (ANS) to extract useful features. Receptive field (rather than fully connected) processing units are used to encode subtle features among their activation patterns. This work initiates ANS applications in the tactile domain and reveals basic characteristics of BEP networks to highly constrained training data.
Cite
Text
Thint and Wang. "Feature Extraction and Clustering of Tactile Impressions with Connectionist Models." International Conference on Machine Learning, 1990. doi:10.1016/B978-1-55860-141-3.50034-1Markdown
[Thint and Wang. "Feature Extraction and Clustering of Tactile Impressions with Connectionist Models." International Conference on Machine Learning, 1990.](https://mlanthology.org/icml/1990/thint1990icml-feature/) doi:10.1016/B978-1-55860-141-3.50034-1BibTeX
@inproceedings{thint1990icml-feature,
title = {{Feature Extraction and Clustering of Tactile Impressions with Connectionist Models}},
author = {Thint, Marcus and Wang, Paul P.},
booktitle = {International Conference on Machine Learning},
year = {1990},
pages = {253-258},
doi = {10.1016/B978-1-55860-141-3.50034-1},
url = {https://mlanthology.org/icml/1990/thint1990icml-feature/}
}