2018 Fiscal Year Final Research Report
Developing Generative Machine Learning and its Digital Circuit Implementation by Leveraging Neuronal Stochastic Behavior
| Project/Area Number |
16K12487
|
|---|
| Research Category |
Grant-in-Aid for Challenging Exploratory Research
|
| Allocation Type | Multi-year Fund |
|---|
| Research Field |
Intelligent informatics
|
|---|
| Research Institution | Kobe University |
|---|
Principal Investigator |
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
上原 邦昭 神戸大学, システム情報学研究科, 教授 (60160206)
|
|---|
| Project Period (FY) |
2016-04-01 – 2019-03-31
|
| Keywords | ニューラルネットワーク / ゆらぎ / 機械学習 |
|---|
| Outline of Final Research Achievements |
This study aimed at developing a new generative machine learning algorithm and implementation method by leveraging the stochastic behavior (uncertainty) and spike-time coding, which biological neural networks have and artificial ones do not. We derived a mathematical model bridging the gap between stochasticity and homeostasis of neurons. We proposed a biologically-plausible learning algorithm by considering the discrete spikes as sampled drawn from a probabilistic distribution. For circuit implementation, we proposed a new approximation method that requires only one-third circuit resources. We also proposed some practical methods based on the stochasticity.
|
| Free Research Field |
機械学習と計算論的神経科額
|
| Academic Significance and Societal Importance of the Research Achievements |
本研究は(1)生物の脳が学習するメカニズムにおいて，従来不明であったゆらぎの貢献や時間的な適応について，数式によるモデルを構築できた．(2)確率的な現象をモデル化できる機械学習手法であるボルツマンマシンを，高密度に電子回路実装する手法を開発した．(3)いわゆる深層学習に確率的要素(ゆらぎや不確実性)を持ち込むことで，高い精度を達成したり，小規模データへ適応可能な手法を開発できた．
|
