Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 55 of the 55 references it cites were found in a registry or an index.

Optogenetics inspired transition metal dichalcogenide neuristors for in-memory deep recurrent neural networks

Open access

Abstract

Shallow feed-forward networks are incapable of addressing complex tasks such as natural language processing that require learning of temporal signals. To address these requirements, we need deep neuromorphic architectures with recurrent connections such as deep recurrent neural networks. However, the training of such networks demand very high precision of weights, excellent conductance linearity and low write-noise- not satisfied by current memristive implementations. Inspired from optogenetics, here we report a neuromorphic computing platform comprised of photo-excitable neuristors capable of in-memory computations across 980 addressable states with a high signal-to-noise ratio of 77. The large linear dynamic range, low write noise and selective excitability allows high fidelity opto-electronic transfer of weights with a two-shot write scheme, while electrical in-memory inference provides energy efficiency. This method enables implementing a memristive deep recurrent neural network with twelve trainable layers with more than a million parameters to recognize spoken commands with >90% accuracy.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41467-020-16985-0

Cites

55 of 55 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Optogenetics-inspired light-driven computational circuits. a Inspiration from optogenetics. Our approach of using PENs in an artificial neural network gives us the flexibility to selectively activate either neurons or synapses. b An I&F circuit is designed using two operational amplifiers and the PEN functionally mimicking optically active ion-channels in optogenetics. Without light activation, the neuron fires at a baseline rate set by the reversal potential Vin of the artificial ion-channel. On illumination by blue light (λ = 445 nm, amplitude = 5 mWcm−2 and pulse width = 10 ms), the conductance of the PEN (Gnr) increases, resulting in a higher frequency of neuronal firing (top). An I&F neuron constructed using conventional resistors does not show any change in firing rate on optical stimulation (bottom). c Photo-modulated spike-timing-dependent plasticity learning rules in PENs. The graphs (read row-wise) depict the modulations of STDP upon illumination with red (623 nm), green (525 nm) and blue (445 nm) light with an intensity = 230 mWcm−2.
Fig. 1 Optogenetics-inspired light-driven computational circuits. a Inspiration from optogenetics. Our approach of using PENs in an artificial neural network gives us the flexibility to selectively activate either neurons or synapses. b An I&F circuit is designed using two operational amplifiers and the PEN functionally mimicking optically active ion-channels in optogenetics. Without light activation, the neuron fires at a baseline rate set by the reversal potential Vin of the artificial ion-channel. On illumination by blue light (λ = 445 nm, amplitude = 5 mWcm−2 and pulse width = 10 ms), the conductance of the PEN (Gnr) increases, resulting in a higher frequency of neuronal firing (top). An I&F neuron constructed using conventional resistors does not show any change in firing rate on optical stimulation (bottom). c Photo-modulated spike-timing-dependent plasticity learning rules in PENs. The graphs (read row-wise) depict the modulations of STDP upon illumination with red (623 nm), green (525 nm) and blue (445 nm) light with an intensity = 230 mWcm−2.CC BY · PubMed Central
Fig. 2 Spatiotemporally selective perturbation. a A 36-element array of PENs was assembled and initially programmed to a common low conductance state (1 nS). Illumination with optical sinusoidal stimuli (red and blue) selectively activated the PENs with each device depicting a non-volatile weight change corresponding to the wavelength of illumination. Blue illumination triggered significant weight changes in the devices resulting in a final spatial conductance map ~102 nS, crossing the pre-set activation/firing threshold of 80 nS. On the other hand, red wavelength stimulation triggered weight changes to a final value of ~13 nS, inactivating these synapses. b A 36-element array comprising of 16 light-sensitive ReS2 and 20 insensitive IWO synapses was assembled and initially programmed to a common low conductance state (5 nS). Spatial conductance maps of the synapses when subjected to red (λ = 623 nm) optical stimuli (intensity = 65 mWcm−2) is shown. Illumination with optical stimuli selectively activated the light-sensitive low-band-gap ReS2 PENs without affecting the light-insensitive high-band-gap IWO synapses. A spatial conductance map readout of the array resembled the image of “N” with the light-sensitive ReS2 PENs spatially arranged accordingly as shown.
Fig. 2 Spatiotemporally selective perturbation. a A 36-element array of PENs was assembled and initially programmed to a common low conductance state (1 nS). Illumination with optical sinusoidal stimuli (red and blue) selectively activated the PENs with each device depicting a non-volatile weight change corresponding to the wavelength of illumination. Blue illumination triggered significant weight changes in the devices resulting in a final spatial conductance map ~102 nS, crossing the pre-set activation/firing threshold of 80 nS. On the other hand, red wavelength stimulation triggered weight changes to a final value of ~13 nS, inactivating these synapses. b A 36-element array comprising of 16 light-sensitive ReS2 and 20 insensitive IWO synapses was assembled and initially programmed to a common low conductance state (5 nS). Spatial conductance maps of the synapses when subjected to red (λ = 623 nm) optical stimuli (intensity = 65 mWcm−2) is shown. Illumination with optical stimuli selectively activated the light-sensitive low-band-gap ReS2 PENs without affecting the light-insensitive high-band-gap IWO synapses. A spatial conductance map readout of the array resembled the image of “N” with the light-sensitive ReS2 PENs spatially arranged accordingly as shown.CC BY · PubMed Central
Fig. 3 Optically addressable multi-level memory for DRNNs. An input optical pulse train of constant pulse width and interval (10 s, 65 mWcm−2) intensity resulted in a precise stepped increase in conductance equivalent to 980 distinct conductance states. a shows a magnified version of the conductance states 1–16 and 965–980 activated by 445-nm wavelength of optical stimuli. To maximise the dynamic range of linear response, the devices were electrically pre-programmed at –40 V to ensure a low initial conductance state. This electrical bias was also maintained constant throughout the optical writing steps. Analysis of endurance during write-erase operations. Application of synergistic optoelectronic pulses result in precisely controlled near-symmetric bidirectional weight changes via optical potentiation and electrical depression as shown in b–d. Positive weight changes activated by 623-nm optical stimuli were erased by electrical pulses of amplitude +20 V. Higher electrical voltages were applied at the back gate to erase the higher degree of potentiation induced by shorter wavelength optical stimulation. For instance, positive weight changes of higher magnitude due to 525/445 nm optical activation were erased by electrical pulses of amplitude +35/40 V respectively. e–j ΔG vs G box and whisker plots – a measure of endurance and write noise during cycling tests. The box portion of the box plot are defined by two lines at the 25th and 75th percentiles and the line drawn inside the box represents the median (50th percentile). Plots e–g depict the variation in the changes in conductance during the optical write operations, while h–j depict the variations during the electrical erase operations for all the three wavelengths of optical stimuli. k Optoelectronic abacus operation. Precise optical potentiation and electrical depression enabled facile emulation of arithmetic operations, analogous to an abacus. Multiple PENs were employed to represent the unit’s (blue bead) and ten’s (red bead) place, and programming steps were designed as per the arithmetic operation under calculation. Optical stimulations resulted in potentiation represented by the rightward sliding of the beads, while electrical stimulations caused depression (leftward sliding of the beads). The methodology of the arithmetic operations are indicated with necessary illustrations as insets.
Fig. 3 Optically addressable multi-level memory for DRNNs. An input optical pulse train of constant pulse width and interval (10 s, 65 mWcm−2) intensity resulted in a precise stepped increase in conductance equivalent to 980 distinct conductance states. a shows a magnified version of the conductance states 1–16 and 965–980 activated by 445-nm wavelength of optical stimuli. To maximise the dynamic range of linear response, the devices were electrically pre-programmed at –40 V to ensure a low initial conductance state. This electrical bias was also maintained constant throughout the optical writing steps. Analysis of endurance during write-erase operations. Application of synergistic optoelectronic pulses result in precisely controlled near-symmetric bidirectional weight changes via optical potentiation and electrical depression as shown in b–d. Positive weight changes activated by 623-nm optical stimuli were erased by electrical pulses of amplitude +20 V. Higher electrical voltages were applied at the back gate to erase the higher degree of potentiation induced by shorter wavelength optical stimulation. For instance, positive weight changes of higher magnitude due to 525/445 nm optical activation were erased by electrical pulses of amplitude +35/40 V respectively. e–j ΔG vs G box and whisker plots – a measure of endurance and write noise during cycling tests. The box portion of the box plot are defined by two lines at the 25th and 75th percentiles and the line drawn inside the box represents the median (50th percentile). Plots e–g depict the variation in the changes in conductance during the optical write operations, while h–j depict the variations during the electrical erase operations for all the three wavelengths of optical stimuli. k Optoelectronic abacus operation. Precise optical potentiation and electrical depression enabled facile emulation of arithmetic operations, analogous to an abacus. Multiple PENs were employed to represent the unit’s (blue bead) and ten’s (red bead) place, and programming steps were designed as per the arithmetic operation under calculation. Optical stimulations resulted in potentiation represented by the rightward sliding of the beads, while electrical stimulations caused depression (leftward sliding of the beads). The methodology of the arithmetic operations are indicated with necessary illustrations as insets.CC BY · PubMed Central
Fig. 4 Simulation of a DRNN for keyword recognition. A deep neural network with 12 trainable layers including convolutional, fully connected, and recurrent LSTM layers is used to classify 12 different spoken digits. a The detailed architecture of the network is shown with all filter sizes and dimensions mentioned. b Highly linear weight update of PENs allow us to transfer high precision weights from offline learnt DRNNs (using sophisticated learning schemes with mini-batch averaging for smooth convergence) along with online learning with blind updates for simpler feed-forward networks. c The accuracy obtained in electrical inference with a two-shot opto-electronic write of weights is within 2.5% of the software simulation with floating-point numbers. The linear dynamic range of the PEN is limited to ~600 due to the two-shot write scheme. Simulating only one layer of convolution, fully connected or LSTM layers show the recurrent LSTM layers to be most sensitive to low weight dynamic range. d PENs used in our work exhibit an order of magnitude higher linear dynamic range than other recent reports, enabling simulation of DRNNs for speech recognition with an order of magnitude higher weights than other reported simulations for handwritten digit recognition.
Fig. 4 Simulation of a DRNN for keyword recognition. A deep neural network with 12 trainable layers including convolutional, fully connected, and recurrent LSTM layers is used to classify 12 different spoken digits. a The detailed architecture of the network is shown with all filter sizes and dimensions mentioned. b Highly linear weight update of PENs allow us to transfer high precision weights from offline learnt DRNNs (using sophisticated learning schemes with mini-batch averaging for smooth convergence) along with online learning with blind updates for simpler feed-forward networks. c The accuracy obtained in electrical inference with a two-shot opto-electronic write of weights is within 2.5% of the software simulation with floating-point numbers. The linear dynamic range of the PEN is limited to ~600 due to the two-shot write scheme. Simulating only one layer of convolution, fully connected or LSTM layers show the recurrent LSTM layers to be most sensitive to low weight dynamic range. d PENs used in our work exhibit an order of magnitude higher linear dynamic range than other recent reports, enabling simulation of DRNNs for speech recognition with an order of magnitude higher weights than other reported simulations for handwritten digit recognition.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 55 references

Research works printed but not linked: 2

  1. Fuller, E. J. et al. Li-ion synaptic transistor for low power analog computing. Adv. Mater. 10.1002/adma.201604310 (2016). · DOI · Paper page
  2. Warden, P. Speech commands: a public dataset for single-word speech recognition. http://download.tensorflow.org/data/speech_commands_v0. Vol. 1 (2017).

Cited by

See all 68

Metadata

Authors
John, R. A., Acharya, J., Zhu, C., Surendran, A., Bose, S. K., Chaturvedi, A., Tiwari, N., Gao, Y., He, Y., Zhang, K. K., Xu, M., Leong, W. L., Liu, Z., Basu, A., & Mathews, N.
Affiliations
Nanyang Technological University (SG)
Corresponding author
  • Basu, Arindam
  • Mathews, Nripan
Venue
Nature Communications
Published
2020-06-25
Volume
11
Issue
1
Pages
3211-3211
ISSN
2041-1723 (ISSN-L)
OpenAlex
W3037369528
Cited by
68
References
55
Citations by year
2026: 6 · 2025: 13 · 2024: 12 · 2023: 11 · 2022: 14 · 2021: 9 · 2020: 3
FWCI
3.18
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • DR-NTU (Nanyang Technological University) (submittedVersion, cc-by)
Funders
Ministry of Education - Singapore (RG 166/16, MOE2015-T2-2-007, MOE2015-T2-2-043, MOE2017-T2-2-136, MOE2016-T2-1-100)
Topics
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Advanced Memory and Neural Computing
  • Physical Sciences › Computer Science › Artificial Intelligence › Neural Networks and Reservoir Computing
  • Physical Sciences › Engineering › Electrical and Electronic Engineering › Ferroelectric and Negative Capacitance Devices
Keywords
transition metal dichalcogenides, neuristor, memristors, in-memory computing, deep recurrent neural networks, signal-to-noise ratio, energy-efficient inference

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Optogenetics inspired transition metal dichalcogenide neuristors for in-memory deep recurrent neural networks. https://app.cochat.ai/papers/10.1038/s41467-020-16985-0
The paper
John, R. A., Acharya, J., Zhu, C., Surendran, A., Bose, S. K., Chaturvedi, A., Tiwari, N., Gao, Y., He, Y., Zhang, K. K., Xu, M., Leong, W. L., Liu, Z., Basu, A., & Mathews, N. (2020). Optogenetics inspired transition metal dichalcogenide neuristors for in-memory deep recurrent neural networks. Nature Communications. https://doi.org/10.1038/s41467-020-16985-0

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.