Spike time dependent plasticity (STDP) is a learning rule in biology, where the time-correlation of narrow pre- and post-synaptic spikes (tspike ≈5ms) is tracked across a wide learning window (LW) of time (tLW = ±80ms) by neurotransmitter dynamics at the synapse such that tLW:tspike> 20. In hardware, resistive random access memory (RRAM) (1M)-based synapse shows STDP by the superposition of long preand post-synaptic neural waveforms comparable to the timescale of the learning window. However, this artificially limits the spike rate and needs a complicated peripheral circuit to generate the waveform, which has an area penalty. In this letter, we propose a PrxCa1-xMnO3 (PCMO)-based RRAM (1M) with an impact-ionization (II)based silicon (Si) NIPIN (n-iv-δ p-i-n) selector (1S) diode as a synapse to operate with only short spikes. The NIPIN device transient response is utilized as a clock at the synapse to compute a preand post-spike time correlation, such that the 1M1S synapse requires short pulses instead of long waveforms. We experimentally demonstrate that very short (tspike ~ 80 ns) square pulses are required to generate STDP while the learning window is tLW ≈ ± 1.5 μs, which enables tLW:tspike ≈19. Furthermore, a hardware acceleration of 1000× over biology is shown. The square pulse scheme avoids complex waveforms and related complicated circuits. Thus, the synaptic time-keeping is demonstrated that enables biologically realistic SNN for future brain inspired computing.
%0 Journal Article
%1 8703727
%A Das, B.
%A Lele, A.
%A Kumbhare, P.
%A Schulze, J.
%A Ganguly, U.
%D 2019
%J IEEE Electron Device Letters
%K iht j.schulze.iht journal
%N 6
%P 850-853
%R 10.1109/LED.2019.2914406
%T PrxCa1–xMnO3-Based Memory and Si Time-Keeping Selector for Area and Energy Efficient Synapse
%U https://ieeexplore.ieee.org/document/8703727/
%V 40
%X Spike time dependent plasticity (STDP) is a learning rule in biology, where the time-correlation of narrow pre- and post-synaptic spikes (tspike ≈5ms) is tracked across a wide learning window (LW) of time (tLW = ±80ms) by neurotransmitter dynamics at the synapse such that tLW:tspike> 20. In hardware, resistive random access memory (RRAM) (1M)-based synapse shows STDP by the superposition of long preand post-synaptic neural waveforms comparable to the timescale of the learning window. However, this artificially limits the spike rate and needs a complicated peripheral circuit to generate the waveform, which has an area penalty. In this letter, we propose a PrxCa1-xMnO3 (PCMO)-based RRAM (1M) with an impact-ionization (II)based silicon (Si) NIPIN (n-iv-δ p-i-n) selector (1S) diode as a synapse to operate with only short spikes. The NIPIN device transient response is utilized as a clock at the synapse to compute a preand post-spike time correlation, such that the 1M1S synapse requires short pulses instead of long waveforms. We experimentally demonstrate that very short (tspike ~ 80 ns) square pulses are required to generate STDP while the learning window is tLW ≈ ± 1.5 μs, which enables tLW:tspike ≈19. Furthermore, a hardware acceleration of 1000× over biology is shown. The square pulse scheme avoids complex waveforms and related complicated circuits. Thus, the synaptic time-keeping is demonstrated that enables biologically realistic SNN for future brain inspired computing.
@article{8703727,
abstract = {Spike time dependent plasticity (STDP) is a learning rule in biology, where the time-correlation of narrow pre- and post-synaptic spikes (tspike ≈5ms) is tracked across a wide learning window (LW) of time (tLW = ±80ms) by neurotransmitter dynamics at the synapse such that tLW:tspike> 20. In hardware, resistive random access memory (RRAM) (1M)-based synapse shows STDP by the superposition of long preand post-synaptic neural waveforms comparable to the timescale of the learning window. However, this artificially limits the spike rate and needs a complicated peripheral circuit to generate the waveform, which has an area penalty. In this letter, we propose a PrxCa1-xMnO3 (PCMO)-based RRAM (1M) with an impact-ionization (II)based silicon (Si) NIPIN (n-iv-δ p-i-n) selector (1S) diode as a synapse to operate with only short spikes. The NIPIN device transient response is utilized as a clock at the synapse to compute a preand post-spike time correlation, such that the 1M1S synapse requires short pulses instead of long waveforms. We experimentally demonstrate that very short (tspike ~ 80 ns) square pulses are required to generate STDP while the learning window is tLW ≈ ± 1.5 μs, which enables tLW:tspike ≈19. Furthermore, a hardware acceleration of 1000× over biology is shown. The square pulse scheme avoids complex waveforms and related complicated circuits. Thus, the synaptic time-keeping is demonstrated that enables biologically realistic SNN for future brain inspired computing.},
added-at = {2019-06-25T13:34:05.000+0200},
author = {{Das}, B. and {Lele}, A. and {Kumbhare}, P. and {Schulze}, J. and {Ganguly}, U.},
biburl = {https://puma.ub.uni-stuttgart.de/bibtex/2f24abd0e61fd5287a56e7965e3b5dbb9/ihtpublikation},
doi = {10.1109/LED.2019.2914406},
interhash = {31f34b506f4304c1ac20c2e87f961450},
intrahash = {f24abd0e61fd5287a56e7965e3b5dbb9},
issn = {0741-3106},
journal = {IEEE Electron Device Letters},
keywords = {iht j.schulze.iht journal},
month = {June},
number = 6,
pages = {850-853},
timestamp = {2019-06-25T11:34:35.000+0200},
title = {PrxCa1–xMnO3-Based Memory and Si Time-Keeping Selector for Area and Energy Efficient Synapse},
url = {https://ieeexplore.ieee.org/document/8703727/},
volume = 40,
year = 2019
}