The 2nd Neuromorphic Research Retreat in AIST

`The 2nd Neuromorphic Research Retreat in AIST:

from Mathematics to Diabetes’

One-day workshop following the previous one <http://tinyurl.com/ybyyb69x> held on 26/10/2016. We offer a meeting point of different communities and stimulate exchanges of a broad spectrum of researchers involved in ``neuromorphic electronics.'' Everybody is welcome!!

Program

10:00–10:10

Introduction

Isao H. Inoue

AIST

(1) 10:10–11:00 (40 mins talk + 10 mins discussion)

Mathematical approach to energy efficient neural information processing.

Gouhei Tanaka

University of Tokyo

Artificial intelligence has been considerably developed in recent years and its performance has outperformed that of human brain in some tasks. However, such an extraordinary ability of artificial intelligence currently requires a huge amount of data, computation, and energy in many cases. For more utilization of artificial intelligence technologies in the real world, it is expected to realize energy efficient artificial intelligence hardware with high performance. In this presentation, I will introduce mathematical approaches to developing energy efficient neural information processing systems. The topics include the reduction of communication costs in recurrent neural networks and the energy efficient implementation of reservoir computing using nonlinear dynamics in physical systems.

(2) 11:00–11:50 (40 mins talk + 10 mins discussion)

Toward the goal of diabetes treatment: theoretical modeling of beta-cells.

Kantaro Fujiwara

Tokyo University of Science

Pancreatic beta-cells are responsible for maintaining blood glucose homeostasis by secreting insulin. However, in some individuals, beta-cells fail to secrete sufficient amounts of insulin, which causes diabetes. For diabetes treatment, understanding insulin secretion mechanism by theoretical modeling is effective. In this study, I construct a model of the electrical activity of beta-cells, focusing on non-selective cation channel, which is recently reported to play an important role for insulin secretion. I would like to discuss the applicability to diabetes treatment by electrically controlling such ionic channels.


11:50–13:30 (lunch)

(3) 13:30–14:20 (40 mins talk + 10 mins discussion)

Spin torque oscillator based artificial neural network

Hiroko Arai

AIST

Artificial neural networks (ANNs) are the key technology for machine learning involving deep learning, which can be applied in the fields such as object detection and recognition. The oscillation model is one of the neural models for artificial neuron, which imitates periodic spiking behaviors of the biological neuron. In order to obtain highly integrated oscillators as processing units for ANNs, the size of the oscillator should desirably be of a nano-order scale. In this study, we propose a processing unit consisting of spin torque oscillators (STOs) with double point contact which provides sigmoid function-like input-output relation. A task-solving demonstration to recognize handwriting digits using the STO-based ANN is also performed.

(4) 14:20–15:10 (40 mins talk + 10 mins discussion)

Simple design of the neuron circuit using phase transition materials.

Takeaki Yajima

University of Tokyo

Neuromorphic computing provides a promising low-energy alternative to the conventional logic- based computing. It is based on the network of neurons and synapses, whose functions have been implemented by analogue circuits. In this talk, I will talk about our recent study on the novel neuron circuits, where the large complexity of the analogue circuits is mitigated by the use of metal-insulator transition materials, VO2. The first half will focus on the basic neuron function called “leaky integrate and fire”, which is simply implemented by the memristive Joule-heat integration and the resultant metal-insulator transition. The later half will discuss more precise dynamics of spiking neurons, and explore how to simply implement it based on the nonlinearity of phase transitions.

15:10–15:30 (break)

(5) 15:30–16:20 (40 mins talk + 10 mins discussion)

SrTiO3-based architectures for deep learning.

Pablo Stoliar

CIC nanoGUNE (Donostia–San Sebastian, Spain), and AIST

We focus on the direct hardware implementation of artificial neural networks. For this purpose, it is necessary to develop both basic building blocks and architectures. Our specific aim is to utilize oxide electronics for the development, because the oxide-based devices present substantial advantages respect to the silicon counterparts. We have already developed artificial synapses and neurons based on SrTiO3, a prototypal oxide material. Now, we are developing necessary circuital architectures that will take advantage of the full potential of these devices. In this talk, we will present our devices and discuss our roadmap to go from single devices to functional neuromorphic circuits.

16:20–16:30

Closing Remarks

Isao H. Inoue

AIST