Potentials and Limitations of Deep Neural Networks for Cognitive Robots

Although Deep Neural Networks reached remark-able performance on several benchmarks and even gained scien-tific publicity, they are not able to address the concept of cognitionas a whole. In this paper, we argue that those architecturesare potentially interesting for cognitive robots regarding theirperceptual representation power for audio and vision data.We identify crucial settings for cognitive robotics where deepneural networks have as yet only contributed little comparedto the challenges in this area. Finally, we highlight the ratherunexplored area of Reservoir Computing for sequence learning.This new paradigm of learning recurrent neural networks in afast and robust way qualifies to be an integral part of cognitiverobots and may inspire novel developments.

@InProceedings{JW17, 
 	 author =  {Jirak, Doreen and Wermter, Stefan},  
 	 title = {Potentials and Limitations of Deep Neural Networks for Cognitive Robots}, 
 	 booktitle = {EUCog Meeting Proceedings},
 	 editors = {},
 	 number = {},
 	 volume = {},
 	 pages = {},
 	 year = {2017},
 	 month = {Nov},
 	 publisher = {EUCog Meeting},
 	 doi = {10.48550/arXiv.1805.00777}, 
 }