Analog Neuron Hierarchy

0507515 - ÚI 2021 RIV GB eng J - Článek v odborném periodiku
Šíma, Jiří
Analog Neuron Hierarchy.
Neural Networks. Roč. 128, August 2020 (2020), s. 199-215. ISSN 0893-6080. E-ISSN 1879-2782
Grant CEP: GA ČR(CZ) GA19-05704S
Institucionální podpora: RVO:67985807
Klíčová slova: recurrent neural network * analog neuron hierarchy * deterministic context-free language * Turing machine * Chomsky hierarchy
Obor OECD: Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)
Impakt faktor: 8.050, rok: 2020
Způsob publikování: Omezený přístup
http://dx.doi.org/10.1016/j.neunet.2020.05.006

In order to refine the analysis of the computational power of discrete-time recurrent neural networks (NNs) between the binary-state NNs which are equivalent to finite automata (level 3 in the Chomsky hierarchy), and the analog-state NNs with rational weights which are Turing complete (Chomsky level 0), we study an intermediate model alphaANN of a binary-state NN that is extended with alpha >= 0 extra analog-state neurons. For rational weights, we establish an analog neuron hierarchy 0ANNs subset 1ANNs subset 2ANNs subseteq 3ANNs and separate its first two levels. In particular, 0ANNs coincide with the binary-state NNs (Chomsky level 3) being a proper subset of 1ANNs which accept at most context-sensitive languages (Chomsky level 1) including some non-context-free ones (above Chomsky level 2). We prove that the deterministic (context-free) language L_# = { 0^n1^n | n >= 1 } cannot be recognized by any 1ANN even with real weights. In contrast, we show that deterministic pushdown automata accepting deterministic languages can be simulated by 2ANNs with rational weights, which thus constitute a proper superset of 1ANNs. Finally, we prove that the analog neuron hierarchy collapses to 3ANNs by showing that any Turing machine can be simulated by a 3ANN having rational weights, with linear-time overhead.
Trvalý link: http://hdl.handle.net/11104/0298502