1998 Fiscal Year Final Research Report Summary
Implementation of Ultra Micro Boltzmann Machine Neuron
| Project/Area Number |
09650375
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
電子デバイス・機器工学
|
|---|
| Research Institution | HOKKAIDO UNIVERSITY |
|---|
Principal Investigator |
AKAZAWA Masamichi Grad.School of Eng., Hokkaido Univ., Asso.Pro., 大学院・工学研究科, 助教授 (30212400)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
WU Nan-Jian Faculty of Electro-Communications, The University of Electro-Communications, Ass, 電気通信学部, 助教授 (00250481)
AMEMIYA Yoshihito Grad.School of Eng., Hokkaido Univ., Pro., 大学院・工学研究科, 教授 (80250489)
|
|---|
| Project Period (FY) |
1997 – 1998
|
| Keywords | Single-Electron Tunneling / Neural Network / Boltzmann Machine / Hopfield Network / Neuron |
|---|
| Research Abstract |
We have investigated ways of using a single-electron-tunneling (SET) circuit to solve problems in neural networks, and have obtained the following results :
1) A simple circuit for a Boltzmann machine neuron : The circuit for the Boltzmann machine neuron that has a stochastic response is usually complicated. We found, however, that by using a single-electron-tunneling circuit, we can construct a compact neuron circuit in which the Coulomb blockade condition is adjusted appropriately, It can produce an output of a random 1-0 bit stream with the probability for an output of 1 controlled by an input signal. A network consisting of more than two neurons can be constructed. So called 'annealing' operation of the network is available by controlling the bias voltage.
2) A way to eliminate the local-minimum problem : If we utilize the cotunneling phenomenon found in single-electron circuits, we can obtain a neural network without the local-minimum problem. In this network, two or more threshold elements can change their outputs simultaneously in a coherent combination. Therefore a state transition with a large Hamming distance can occur and the local-minimum difficulty can disappear. As a result, the global-minimum energy state can always be achieved. Use of this property makes possible the development of novel computation devices that solve combinatorial problems without hindrance from the local-minimum difficulty.
|
