Development of Human-timescale Neural Circuits using Emerging Neuromorphic Devices

2018 Fiscal Year Annual Research Report

• Project/Area Number: 18H05911
• Allocation Type: Single-year Grants
• Research Institution: National Institute of Advanced Industrial Science and Technology
• Principal Investigator: Stoliar Pablo 国立研究開発法人産業技術総合研究所, エレクトロニクス・製造領域, 主任研究員 (40824545)
• Project Period (FY): 2018-08-24 – 2020-03-31
• Keywords: Neuromorphic systems / Emerging devices / Recurrent neural net. / Homeostasis mechanisms / Ferroelec. tunnel junct.

Outline of Annual Research Achievements

So far, as planned, I developed an artificial neuron using ferroelectric tunnel junctions and a set of tools for studying its behaviour in recurrent neural networks (RNN). These neurons are based on a physical mechanism radically different than conventional electronics: resistive switching by controlling the polarization reversal in the ferroelectric barrier. It makes them much smaller and energy efficient to interact with human-timescale signals.
For developing the neurons, I designed and build ad-hoc electric instrumentation required to characterize the devices with neuromorphic signals. Then, I developed a concentrated parameters model, and finally, the whole circuit. They integrate, which is a crucial function in neurons, by gradually switching the ferroelectric barrier.
I also developed a (simulated) RNN, based on a pool of 1000 neurons with sparse connectivity. Before training, the system behaves chaotic. I then implemented a training algorithm that allows it to reproduce arbitrary sequences. For the training, the synaptic connections to a selected output neuron are tune by supervised learning. I assembled a computer system for doing simulations requiring both intense GPU and CPU computations, which was necessary for training.
In the second part of the project, based on these neurons, I will study the stability of RNN’s, develop architectures to reduce energy requirements, and develop instrumentation to implement and characterize neuromorphic systems.

Current Status of Research Progress

2: Research has progressed on the whole more than it was originally planned.

Reason

I proposed 3 types of research, namely:
1) Fabrication, modelling and characterization of ferroelectric tunnel junctions (FTJ).
2) Design of a neuron using FTJ.
3) Design an RNN using the FTJ-based neurons, including software for training and inference.
I completed 100% of these 3 researchers.

Strategy for Future Research Activity

For the promotion of the result, I am preparing the following set of publication base on the results: 1) Long time-scale neurons based on emerging research devices (completed 25%). 2) FTJ-based STDP synapses (draft almost ready). 3) Ultracompact LIF neuron (draft submitted). 4) Source-measuring unit for characterizing FTJ’s (draft ready to submit). 5) Architectures to implement convolutional layers with resistive switching devices (draft submitted).