CoChat

Face classification using electronic synapses

Open access

The DOI matches its Crossref record; 33 of the 34 references it cites were found in a registry or an index.

Abstract

Conventional hardware platforms consume huge amount of energy for cognitive learning due to the data movement between the processor and the off-chip memory. Brain-inspired device technologies using analogue weight storage allow to complete cognitive tasks more efficiently. Here we present an analogue non-volatile resistive memory (an electronic synapse) with foundry friendly materials. The device shows bidirectional continuous weight modulation behaviour. Grey-scale face classification is experimentally demonstrated using an integrated 1024-cell array with parallel online training. The energy consumption within the analogue synapses for each iteration is 1,000 × (20 ×) lower compared to an implementation using Intel Xeon Phi processor with off-chip memory (with hypothetical on-chip digital resistive random access memory). The accuracy on test sets is close to the result using a central processing unit. These experimental results consolidate the feasibility of analogue synaptic array and pave the way toward building an energy efficient and large-scale neuromorphic system.

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/ncomms15199

Cites

33 of 33 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure 1 The 1T1R architecture and the 1024-cell-1T1R array. (a) Mapping of a one-layer neural network on the 1T1R array, that is, the input of preneuron layer, adaptable synaptic weight and weighted sum output of postneuron layer maps to the pulse input from BL, cell conductance and current output through SL, respectively. In 1T1R, ‘T' represents transistor, ‘R' represents RRAM. (b) The micrograph of a fabricated 1024-cell-1T1R array using fully CMOS compatible fabrication process.
Figure 1 The 1T1R architecture and the 1024-cell-1T1R array. (a) Mapping of a one-layer neural network on the 1T1R array, that is, the input of preneuron layer, adaptable synaptic weight and weighted sum output of postneuron layer maps to the pulse input from BL, cell conductance and current output through SL, respectively. In 1T1R, ‘T' represents transistor, ‘R' represents RRAM. (b) The micrograph of a fabricated 1024-cell-1T1R array using fully CMOS compatible fabrication process.CC BY · PubMed Central
Figure 2 Flowchart of the perceptron model. (a) The training process flow chart. In this demonstration, a batch learning model is used to accelerate the converging speed. Here ‘n' represents the number of pattern, ranging from 1 to 9, ‘i' implies the index of a pixel of an input pattern and can be defined from 1 to 320, ‘j' is the number of output neuron that is 1–3. A correct classification during the inference phase means the active function value of a matching class of the input pattern is greater than other two classes. This network converges when all training patterns are correctly recognized. (b) The schematic of parallel read operation and how a pattern is mapped to the input. (c) The nine training images, which is a cropped and subsampled subset of the Yale Face Database19.
Figure 2 Flowchart of the perceptron model. (a) The training process flow chart. In this demonstration, a batch learning model is used to accelerate the converging speed. Here ‘n' represents the number of pattern, ranging from 1 to 9, ‘i' implies the index of a pixel of an input pattern and can be defined from 1 to 320, ‘j' is the number of output neuron that is 1–3. A correct classification during the inference phase means the active function value of a matching class of the input pattern is greater than other two classes. This network converges when all training patterns are correctly recognized. (b) The schematic of parallel read operation and how a pattern is mapped to the input. (c) The nine training images, which is a cropped and subsampled subset of the Yale Face Database19.CC BY · PubMed Central
Figure 3 The performance of the optimized device and two examples for programming. (a) Typical I–V curve of a single 1T1R cell for a quasi-DC sweep, the gate voltage is 1.8 and 8 V during SET and RESET process, respectively. Inset is a transmission electron microscope (TEM) image of the RRAM device. (b) An example of the typical continuous conductance tuning performance under an identical pulse train condition during SET process, along with the fitting curve. Vwl=2.4 V, Vbl=2.0 V (50 ns), Vsl=0 V. (c) Tuning performance during RESET operation, along with the fitting curve. Vwl=8 V, Vbl=0 V, Vsl=2.3 V (50 ns). (d) An example of the SET programming waveform applied on the first row to adjust the weight, following write-verify scheme. (e) Waveforms for programming without write-verify. (f) The precision measurement result during RESET process using verified pulse train with different amplitudes. y Axis represents the final conductance accuracy (the difference between the target conductance and the measured conductance over the target conductance) after programming from a same initial state 40 μS. (g) y Axis represents the number of pulses needed to reach the target conductance from the same initial state 40 μS. These curves show the relationship of tuning speed with respect to different programming pulse amplitudes.
Figure 3 The performance of the optimized device and two examples for programming. (a) Typical I–V curve of a single 1T1R cell for a quasi-DC sweep, the gate voltage is 1.8 and 8 V during SET and RESET process, respectively. Inset is a transmission electron microscope (TEM) image of the RRAM device. (b) An example of the typical continuous conductance tuning performance under an identical pulse train condition during SET process, along with the fitting curve. Vwl=2.4 V, Vbl=2.0 V (50 ns), Vsl=0 V. (c) Tuning performance during RESET operation, along with the fitting curve. Vwl=8 V, Vbl=0 V, Vsl=2.3 V (50 ns). (d) An example of the SET programming waveform applied on the first row to adjust the weight, following write-verify scheme. (e) Waveforms for programming without write-verify. (f) The precision measurement result during RESET process using verified pulse train with different amplitudes. y Axis represents the final conductance accuracy (the difference between the target conductance and the measured conductance over the target conductance) after programming from a same initial state 40 μS. (g) y Axis represents the number of pulses needed to reach the target conductance from the same initial state 40 μS. These curves show the relationship of tuning speed with respect to different programming pulse amplitudes.CC BY · PubMed Central
Figure 4 The training process of the experimental demonstration. (a) The activation function output value of the first class versus the iteration number using the write-verify scheme. The inset figure zooms in the several last steps. (b) The training process for programming without write-verify. (c) The initial and final conductance distribution comparison of the first row when updating with write-verify. Inset shows the final conductance map. (d) The conductance distribution of the first row and the conductance map for the case without write-verify.
Figure 4 The training process of the experimental demonstration. (a) The activation function output value of the first class versus the iteration number using the write-verify scheme. The inset figure zooms in the several last steps. (b) The training process for programming without write-verify. (c) The initial and final conductance distribution comparison of the first row when updating with write-verify. Inset shows the final conductance map. (d) The conductance distribution of the first row and the conductance map for the case without write-verify.CC BY · PubMed Central
Figure 5 The test result with latency and energy comparison. (a) A standard example of the constructed test patterns with 100 noise pixels with respect to each training image in Fig. 2c. (b) The recognition rate curve of two programming strategies during the test. The x axis represents the amount of noise. (c) The comparison between with (W/) and without (W/O) schemes in terms of latency and energy consumption during training process and testing recognition rate in the normalized format.
Figure 5 The test result with latency and energy comparison. (a) A standard example of the constructed test patterns with 100 noise pixels with respect to each training image in Fig. 2c. (b) The recognition rate curve of two programming strategies during the test. The x axis represents the amount of noise. (c) The comparison between with (W/) and without (W/O) schemes in terms of latency and energy consumption during training process and testing recognition rate in the normalized format.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 33 references

Research works printed but not linked: 6

  1. Le, Q. V. in Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE International Conference 8595–8598 (IEEE, 2013).
  2. Krizhevsky, A., Sutskever, I. & Hinton, G. E. in Advances in Neural Information Processing Systems 1097–1105 (Curran Associates, Inc., 2012).
  3. Fackenthal, R. et al. in Solid-State Circuits Conference Digest of Technical Papers (ISSCC), 2014 IEEE International, 338–339 (IEEE, 2014).
  4. Govoreanu, B. et al. in Electron Devices Meeting (IEDM), 2011 IEEE International, 31.36. 31–31.36. 34 (IEEE, 2011).
  5. Sekar, D. et al. in 2014 IEEE International Electron Devices Meeting, 28.23. 21–28.23. 24 (IEEE, 2014).
  6. Burr, G. et al. in 2015 IEEE International Electron Devices Meeting (IEDM), 4.4. 1–4.4. 4 (IEEE, 2015).

Cited by

Items 201–300 of 918

Metadata

Authors
Yao, P., Wu, H., Gao, B., Eryilmaz, S. B., Huang, X., Zhang, W., Zhang, Q., Deng, N., Shi, L., Wong, H.-S. P., & Qian, H.
Affiliations
  • Tsinghua University (CN)
  • Stanford SystemX Alliance (US)
Corresponding author
Wu, Huaqiang
Venue
Nature Communications
Published
2017-05-12
Volume
8
Issue
1
Pages
15199-15199
ISSN
2041-1723 (ISSN-L)
OpenAlex
W2613205562
Cited by
909
References
34
Citations by year
2026: 41 · 2025: 73 · 2024: 91 · 2023: 106 · 2022: 120 · 2021: 127 · 2020: 132 · 2019: 113 · 2018: 91 · 2017: 14
FWCI
39.80
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (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)
Funders
  • National Science Foundation (1317470)
  • Yale University
  • Arizona State University
  • Stanford SystemX Alliance
  • National Natural Science Foundation of China (61674089)
  • Beijing Innovation Center for Future Chip
  • National Key Research and Development Program of China
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
electronic synapses, resistive random access memory, neuromorphic hardware, face classification, synaptic arrays, energy efficiency, neuromorphic computing

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

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.