Spikes as Regularizers
Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review
Standard
Spikes as Regularizers. / Søgaard, Anders.
ESANN 2017 - Proceedings: 25th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning. ESANN , 2017. p. 371-376 (arXiv.org).Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review
Harvard
APA
Vancouver
Author
Bibtex
}
RIS
TY - GEN
T1 - Spikes as Regularizers
AU - Søgaard, Anders
PY - 2017
Y1 - 2017
N2 - We present a confidence-based single-layer feed-forward learning algorithm SPIRAL(Spike Regularized Adaptive Learning) relying on an encoding of activationspikes. We adaptively update a weight vector relying on confidence estimates andactivation offsets relative to previous activity. We regularize updates proportionallyto item-level confidence and weight-specific support, loosely inspired by the observationfrom neurophysiology that high spike rates are sometimes accompaniedby low temporal precision. Our experiments suggest that the new learning algorithmSPIRAL is more robust and less prone to overfitting than both the averagedperceptron and AROW.
AB - We present a confidence-based single-layer feed-forward learning algorithm SPIRAL(Spike Regularized Adaptive Learning) relying on an encoding of activationspikes. We adaptively update a weight vector relying on confidence estimates andactivation offsets relative to previous activity. We regularize updates proportionallyto item-level confidence and weight-specific support, loosely inspired by the observationfrom neurophysiology that high spike rates are sometimes accompaniedby low temporal precision. Our experiments suggest that the new learning algorithmSPIRAL is more robust and less prone to overfitting than both the averagedperceptron and AROW.
M3 - Article in proceedings
SN - 978-287587039-1
T3 - arXiv.org
SP - 371
EP - 376
BT - ESANN 2017 - Proceedings
PB - ESANN
T2 - 25th European Symposium on Artificial Neural Networks, omputational Intelligence and Machine Learning
Y2 - 26 April 2017 through 28 April 2017
ER -
ID: 195004442