Contrastive auto-encoder for phoneme recognition

Abstract

Speech data typically contains task irrelevant information lying within features. Specifically, phonetic information, speaker characteristic information, emotional information and noise are always mixed together and tend to impair one another for certain task. We propose a new type of auto-encoder for feature learning called contrastive auto-encoder. Unlike other variants of auto-encoders, contrastive auto-encoder is able to leverage class labels in constructing its representation layer. We achieve this by modeling two autoencoders together and making their differences contribute to the total loss function. The transformation built with contrastive auto-encoder can be seen as a task-specific and invariant feature learner. Our experiments on TIMIT clearly show the superiority of the feature extracted from contrastive auto-encoder over original acoustic feature, feature extracted from deep auto-encoder, and feature extracted from a model that contrastive auto-encoder originates from.

A comparison of neural network methods for unsupervised representation learning on the zero resource speech challenge

“…We also introduce an improved cAE architecture and training method that reduces the number of hyperparameters to be tuned, and show that narrow architectures work better, with reduced error rates on a zero-resource language after tuning on English. Unlike similar previous work [9,10,11] our cAE-based systems are fully unsupervised, train on individual frames without context, and use a loss function in the input vector space instead of the representation vector space.…”

Are you eligible? Predicting adulthood from face images via Class Specific Mean Autoencoder

“…No Rifai et al (2011a) Higher order Contractive autoencoder: CAE + Hessian of the output wrt the input. No Zheng et al (2014) Contrastive autoencoder: A term to reduce the intra-class variations between the learned representation of samples belonging to the same class is added at the final layer.…”

“…Contrastive Autoencoder (CsAE) proposed by Zheng et al (2014), is another variant of supervised autoencoder which uses the class label information during training. The loss function of the model is the difference between the output of two sub-autoencoders trained simultaneously on samples belonging to the same class, along with the loss function of each subautoencoder.…”

Class representative autoencoder for low resolution multi-spectral gender classification

“…Recently, supervised extensions of traditional unsupervised model of the autoencoder have also been proposed [17], [18], [19], [20]. Most of these algorithms incorporate class information at the time of feature extraction with an aim to reduce only the intra-class variations.…”