基于MAML方法的佤语孤立词分类  

MAML-based classification of isolated words of the Wa language

作　　者：王翠 王璐 解雪琴 杨建香 和丽华 侯俊龙 潘文林[1] WANG Cui;WANG Lu;XIE Xue-qin;YANG Jian-xiang;HE Li-hua;HOU Jun-long;PAN Wen-lin(School of Mathematics and Computer Science,Yunnan Minzu University,Kunming 650500,China)

机构地区：[1]云南民族大学数学与计算机科学学院,云南昆明650500

出　　处：《云南民族大学学报（自然科学版）》2020年第4期376-381,395,共7页Journal of Yunnan Minzu University:Natural Sciences Edition

基　　金：国家自然科学基金(61866040);云南省教育厅科学研究基金(2019Y0219).

摘　　要：深度神经网络模型直接用于识别有限的佤语语音很容易陷入过拟合的窘境.而元学习善于解决小样本因数据不足导致的过拟合学习问题,利用平摊机制将以往经验应用于新任务学习能够有效减少对目标数据集的要求.选用原始的与模型无关的元学习(MAML)和近似一阶MAML梯度更新的Reptile对佤语语谱图进行识别研究.基于两组对比实验结果表明,元学习具有快速学习能力,并且能显著提高网络的收敛能力和泛化能力.在相同的实验设置条件下,MAML和Reptile对5-way 1-shot的实验准确率分别达到74.5%和61.6%,对5-way 5-shot的识别准确率分别达到94.5%和93.6%.It is easy to fall into the dilemma of overfitting if the deep neural network model is directly used to recognize the limited speech sounds of the Wa language. However, meta-learning is good at solving the overfitting problem caused by insufficient data in few-shot learning, and using the amortization mechanism to apply previous experience to new task learning can effectively reduce the requirements on the target data set. In this paper, the original model-agnostic meta-learning(MAML) and Reptile with approximate first-order MAML gradient update are used to identify the spectrogram of the Wa language. Based on the experimental results of two groups, it is shown that meta-learning has fast learning ability which can significantly improve the convergence ability and generalization ability of the network. In the same experimental conditions, the accuracy rates of MAML and Reptile on 5-way 1-shot are respectively 74.5% and 61.6%, while on 5-way 5-shot are respectively 94.5% and 93.6%.

关 键 词：MAML模型 小样本学习 佤语语谱图 孤立词 

分 类 号：TN912.34[电子电信—通信与信息系统]