Nonlinear Operation of Static-Binary Neuron Circuits and Dynamic Memristive Devices for STDP Learning

Marukame, Takao; Sugino, Junichi; Kitamura, Toshimitsu; Ishikawa, Kazuo; Takahashi, Koji; Tsutsumi, Yutaka; Berdan, Radu; Nomura, Kumiko; Nishi, Yoshifumi

doi:10.1109/iscas.2019.8702643

Cited by 4 publications

(13 citation statements)

References 32 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The observed I-V characteristics did not show abrupt switching but rather a moderate change in the voltage region around −1 and +2 V in the hysteresis curve, providing direct evidence that analog memristive devices are feasible with WO x /MgO. 26) 2.2. Spike-timing-dependent plasticity Biological synapse conductance is changed nonlinearly by a spike change.…”

Section: Memristive Characteristicsmentioning

confidence: 72%

“…To achieve practical inference hardware, the bit precision (binary or continuous) of the weight values has to be taken into account in modeling to design NN hardware. 25,26) Here, we analyze a NN for practical uses on the basis of the most standard topologies and learning algorithms [Fig. 7(a)].…”

Section: Nn Simulations Of Analog Neurons With Resistive Synapsesmentioning

confidence: 99%

“…[5][6][7] Before discussing NN inference circuits, we tried to simplify the learning sequence of the NN by focusing on backward connections between the hidden layer and the output layer. 26) A simple mechanism, called "random backpropagation NN," has been presented. It assigns random synaptic weights for backward processing from output to hidden/input layers.…”

Section: Nn Simulations Of Analog Neurons With Resistive Synapsesmentioning

confidence: 99%

“…It assigns random synaptic weights for backward processing from output to hidden/input layers. 26,30,31) This mechanism can transmit error signals across multiple layers of neurons and performs as effectively as standard backpropagation. 26) Accordingly, the robustness was analyzed with randomized weight values in the connections between the hidden and output layers.…”

Section: Nn Simulations Of Analog Neurons With Resistive Synapsesmentioning

confidence: 99%

“…Annealing in an O 2 atmosphere around 400 °C was carried out to form WO x after film deposition of W. For comparison, we also prepared a similar device structure without MgO film, although we exploited a well-tuned MgO film as a tunnel barrier 23,24) similar to the AlO x /TiO x tunnel barrier used in our previous study. 25) Figure 2 shows transmission electron microscopy images revealing that the interface of the WO x /MgO films was abrupt compared with that of the WO x /TiN film, and that the WO x was polycrystalline, while the MgO was amorphous. Junction areas between 1 and 100 μm 2 were formed by lithography and dry etching.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Integrated analog neurons inspired by mimicking synapses with metal-oxide memristive devices

Marukame

Sugino

Kitamura

et al. 2020

Jpn. J. Appl. Phys.

Self Cite

View full text Add to dashboard Cite

Artificial neural networks (NNs) integrated with emerging devices are one promising approach to overcoming the limitations of artificial intelligence hardware based on existing CMOS technologies. Two-terminal memristive devices consisting of metal-oxide WOx/MgO were fabricated and investigated in order to add synaptic memory functions to NN hardware. The device showed analog conductance changes similar to a biological synapse’s spike-timing dependent plasticity as well as its binarized characteristics, which depend on waveforms of voltage spikes. To exploit such properties of the memristive devices for practical hardware, NN simulations were executed taking into account analog and binary synapses, resulting in good accuracy. For rapid NN prototyping with resistive synapses, we developed an analog-to-digital mixed circuit with variable resistors, which was fabricated using a standard CMOS process. Thirty-two analog neurons, where 1 neuron comprises 32 synapses, performed 1k synaptic operations at approximately 10 mW and generate neuronal firing at approximately less than 2 μs.

show abstract

Section: Memristive Characteristicsmentioning

confidence: 72%

Section: Nn Simulations Of Analog Neurons With Resistive Synapsesmentioning

confidence: 99%