Towards a Brain-inspired Developmental Neural Network by Adaptive Synaptic Pruning | |
Zhao Feifei1,3; Zhang Tielin1; Zeng Yi1,2,3; Xu Bo1,2,3 | |
2017 | |
会议名称 | The 24th international Conference on Neural Information Processing (ICONIP 2017) |
会议日期 | November 14-18 |
会议地点 | Guangzhou, China, 2017 |
摘要 | It is widely accepted that appropriate network topology should be empirically predefined before training a specific neural network learning task. However, in most cases, these carefully designed networks are easily falling into two kinds of dilemmas: 1) When the data is not enough to train the network well, it will get an underfitting result. 2)When networks have learned too much patterns, they are likely to lead to an overfitting result and have a poor performance on processing new data or transferring to other tasks. Inspired by the synaptic pruning characteristics of the human brain, we propose a brain-inspired developmental neural network (BDNN) algorithm by adaptive synaptic pruning (BDNN-sp) which could get rid of the overfitting and underfitting. The BDNN-sp algorithm adaptively modulates network topology by pruning useless neurons dynamically. In addition, the evolutional optimization method makes the network stop on an appropriate network topology with the best consideration of accuracy and adaptability. Experimental results indicate that the proposed algorithm could automatically find the optimal network topology and the network complexity could adaptively increase along with the increase of task complexity. Compared to the traditional topology-predefined networks, trained BDNN-sp has the similar accuracy but better transfer learning abilities. |
关键词 | Brain-inspired Developmental Neural Network Brain-inspired Pruning Rules Structural Plasticity Network Adaptability Synaptic Pruning |
文献类型 | 会议论文 |
条目标识符 | http://ir.ia.ac.cn/handle/173211/23557 |
专题 | 类脑智能研究中心_类脑认知计算 |
通讯作者 | Zeng Yi |
作者单位 | 1.Institute of Automation, Chinese Academy of Sciences, Beijing, China 2.Center for Excellence in Brain Science and Intelligence Technology, Chinese Academy of Sciences, Shanghai, China 3.University of Chinese Academy of Sciences, Beijing, China |
第一作者单位 | 中国科学院自动化研究所 |
通讯作者单位 | 中国科学院自动化研究所 |
推荐引用方式 GB/T 7714 | Zhao Feifei,Zhang Tielin,Zeng Yi,et al. Towards a Brain-inspired Developmental Neural Network by Adaptive Synaptic Pruning[C],2017. |
条目包含的文件 | 下载所有文件 | |||||
文件名称/大小 | 文献类型 | 版本类型 | 开放类型 | 使用许可 | ||
Towards a Brain-insp(387KB) | 会议论文 | 开放获取 | CC BY-NC-SA | 浏览 下载 |
除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。
修改评论