Energy-Time Tradeoff in Recurrent Neural Nets

0472477 - ÚI 2017 RIV CH eng C - Conference Paper (international conference)
Šíma, Jiří
Energy-Time Tradeoff in Recurrent Neural Nets.
Artificial Neural Networks. Methods and Applications in Bio-/Neuroinformatics. Cham: Springer, 2015 - (Koprinkova-Hristova, P.; Mladenov, V.; Kasabov, N.), s. 51-62. Springer Series in Bio-/Neuroinformatics, 4. ISBN 978-3-319-09902-6. ISSN 2193-9349.
[ICANN 2013. International Conference on Artificial Neural Networks /23./. Sofia (BG), 10.09.2013-13.09.2013]
R&D Projects: GA ČR GBP202/12/G061
Institutional support: RVO:67985807
Keywords : energy complexity * recurrent neural network * finite automaton * energy-time tradeoff
Subject RIV: IN - Informatics, Computer Science
https://link.springer.com/content/pdf/bfm:978-3-319-09903-3/1

In this chapter, we deal with the energy complexity of perceptron networks which has been inspired by the fact that the activity of neurons in the brain is quite sparse (with only about 1% of neurons firing). This complexity measure has recently been introduced for feedforward architectures (i.e., threshold circuits). We shortly survey the tradeoff results which relate the energy to other complexity measures such as the size and depth of threshold circuits. We generalize the energy complexity for recurrent architectures which counts the number of simultaneously active neurons at any time instant of a computation. We present our energy-time tradeoff result for the recurrent neural nets which are known to be computationally as powerful as the finite automata. In particular, we show the main ideas of simulating any deterministic finite automaton by a low-energy optimal-size neural network. In addition, we present a lower bound on the energy of such a simulation (within a certain range of time overhead) which implies that the energy demands in a fixedsize network increase exponentially with the frequency of presenting the input bits.
Permanent Link: http://hdl.handle.net/11104/0269779