2021 Fiscal Year Final Research Report
Reservoir-computing based on the field theory of dynamically balanced neuronal networks
Project/Area Number |
19K20359
|
Research Category |
Grant-in-Aid for Early-Career Scientists
|
Allocation Type | Multi-year Fund |
Review Section |
Basic Section 61040:Soft computing-related
|
Research Institution | Nihon University |
Principal Investigator |
|
Project Period (FY) |
2019-04-01 – 2022-03-31
|
Keywords | ニューラルネットワーク / レザボワ計算 / 機械学習 / 脳神経回路 / 統計力学 / 力学系 |
Outline of Final Research Achievements |
We previously constructed a novel statistical mechanical theory for critically balanced neural networks. In this research project, we aimed at improving performance in a machine learning problem called reservoir computing based on this theory. First, we succeeded in developing a response theory that describes the information the neural networks probabilistically read out in response to external inputs. Using this theory, we showed that the way neural networks retain information about the input largely depends on the details of their connectivity. Then, we numerically showed that the performance in reservoir computing can be improved by pretraining balanced neural networks so that the information-retaining properties of the networks become favourable to learning. Furthermore, we numerically demonstrated that generalisation performance of the networks is improved by using an algorithm derived from a learning theory for balanced neural networks.
|
Free Research Field |
生物物理学
|
Academic Significance and Societal Importance of the Research Achievements |
人工神経回路を用いたAIは一部の情報処理では動物脳を超えた性能を示すが、過去の情報を記憶して未来に活かす潜在能力を秘めたリカレント神経回路については未だ技術的なブレイクスルーや理論的な解明が待たれている段階である。研究代表者は神経回路の学習に関して近年一定の成功をおさめている統計力学的アプローチを用い、これまで調べられていなかった動物脳にみられる動的均衡という性質を持つリカレント神経回路の理論を世界に先駆けて樹立し、本研究課題を通してその学習における有用性を明らかにした。この研究結果により、リカレント神経回路による過去の記憶の読み出し性能を飛躍的に向上する手がかりが得られた可能性がある。
|