ART2wNF及其稳定性-可塑性动态平衡性能分析

ART2wNF and Its Stability-plasticity Dynamic Trade-off Analysis

稳定性-可塑性两难问题的核心是系统如何在不削弱或忘记已学习模式的同时,自适应地学习新事物.目前公认自适应谐振理论(Adaptive resonance theory,ART)能够部分解决稳定性–可塑性两难问题,但依然存在学习受样本输入顺序影响大,且存在学习中心渐变样本时,带来的所谓模式漂移的问题.受进化生物学关于人类学习的幼态延续特征的启发,本文为每个F2层节点配备活跃度指示器λ,并将其反馈回F1层参与STM(Short term memory)向量的计算,使这种新型ART2网络在行为特征上具备幼态延续的显著特征,本文称之为ART2wNF(Adaptive resonance theory with neoteny feature).论文从理论上证明算法的可行性,并通过分析对随机生成样本集合的学习过程,对比了ART2wNF算法与常规ART2网络在可塑性、稳定性方面的差异以及ART2wNF在克服样本输入顺序影响等方面的优势.

Stability-plasticity dilemma is how to build a sys- tem that is adaptive enough to learn new things while not dilut- ing/forgetting previously learned patterns. It is well known that ART （adaptive resonance theory） network can partially solve the stability-plasticity dilemma, but the behavior of ART network is uncertain due to the input order of samples and the pattern drift problem which is also notable for patterns with gradually changed center. Inspired by the human neoteny phenomenon discussed in evolution biology, in order to record the stimulating degree, we suggest that each node in F2 layer be accompanied by an activity indicator A, which is also fed back to F1 layer as a parameter of the calculation of STM （short term memory） vectors. The modified ART2 network has the remarkable fea- ture of neoteny during learning process and is called ART2wNF （adaptive resonance theory with neoteny feature） in this paper. The feasibility of the arithmetic is theoretically proved for the introduction of A at first. Then the performance and distinct- ness of ART2wNF in stability and plasticity are compared with ART2 by analyzing the learning process for randomly generated samples. It also shows the distinctive ability to overcome the shortage of ART2 caused by different input orders.