增量式神经网络聚类算法

摘要

神经网络模型具有强大的问题建模能力，但是传统的反向传播算法只能进行批量监督学习，并且训练开销很大。针对传统算法的不足，提出全新的增量式神经网络模型及其聚类算法。该模型基于生物神经学实验证据，引入新的神经元激励函数和突触调节函数，赋予模型以坚实的统计理论基础。在此基础上，提出一种自适应的增量式神经网络聚类算法。算法中引入“胜者得全”式竞争等学习机制，在增量聚类过程中成功避免了“遗忘灾难”问题。在经典数据集上的实验结果表明：该聚类算法与K-means等传统聚类算法效果相当，特别是在增量学习任务的时空开销方面具有较大优势。%Neural network model is powerful in problem modelling.But the traditional back propagating algorithm can only execute batch supervised learning,and its time expense is very high.According to these problems,a novel incremental neural network model and the corresponding clustering algorithm were put forward.This model was supported by biological evidences,and it was built on the foundation of novel neuron’s activation function and the synapse adjusting function.Based on this,an adaptive incremental clustering algorithm was put forward,in which mechanisms such as “winner-take-all”were introduced.As a result,“catastrophic forgetting”problem was successfully solved in the incremental clustering process.Experiment results on classic datasets show that this algorithm ’s performance is comparable with traditional clustering models such as K-means.Especially,its time and space expenses on incremental tasks are much lower than traditional clustering models.