Research On Temporal Alignment Methods With Autoencoder Regularization

Temporal alignment aims at aligning two sequences in time coordinate for minimizing the distance between both sequences by time warping methods.Dynamic time warping(DTW)is a typical time warping method,which has been widely used in machine learning and pattern recognition.However,its representation capacity is still not enough strong to boost alignment performance.For this reason,this thesis is the first attempt to treat autoencoder as the regularization term to enhance the alignment performance of dynamic time warping from two aspects,i.e.,shallow nonlinear representation and deep representation,respectively.The main work of this thesis is as follows:1.A nonlinear AutoEncoder Regularized Canonical Time Warping(AECTW)method is proposed.DTW and its effective variants,especially Canonical Time Warping(CTW),have significantly improved the performance of temporal alignment tasks.However,their linear representation ability is still so limited that alignment resluts are not satisfactory.To address this issue,this thesis proposes a non-linear Autoencoder Regularized Canonical Time Warping(AECTW)method.Specifically,AECTW combines CTW with the nonlinear autoencoder regularization term to learn nonlinear underlying structure within the dataset by explicitly non-linear activation functions,meanwhile enhancing the ability to reconstruct low-dimensional representations.Benefitting from nonlinear representation and reconstruction ability,AECTW can improve the alignment performance of CTW.Experimental results on the synthetic data and the actual activity recognition datasets show that AECTW is superior to the representative time warping methods.2.The stacked temporal alignment(STA)framework is developed.The alignment performance of existing time warping methods depend heavily on the ability of feature representation.Actually,they are shallow in the sense that they employ original data or their once projections to perform temporal alignment tasks.Fortunately,recent advance in deep learning have witnessed that deeper architecture is more beneficial for effective feature representation.In this thesis,we first incorporate temporal alignment model with deep representation,and propose the Stacked Temporal Alignment(STA)framework.The STA framework is more flexible,and offers the promise of integrating effective strategies of current time warping methods into deep architecture,and meanwhile provides a consistent perspective to understand the existing time warping methods.To balance the alignment performance as well as time efficiency,we then explore a Stacked Marginal Time Warping(SMTW)method based on STA.SMTW utilizes marginal stacked autoencoder(mSDA)as the regularization term for robust representation and reveals the nonlinear structure by the layer-wise activation function.Experimental results on synthetic data and human activity recognition datasets show that SMTW achieves the best temporal alignment performance,meanwhile is very competitive in time compared with baseline time warping methods.This also verifies the effectiveness of the proposed STA framework.