Publications

In image classification, visual separability between different object categories is highly uneven, and some categories are more difficult to distinguish than others. Such difficult categories demand more dedicated classifiers. However, existing deep convolutional neural networks (CNN) are trained as flat N-way classifiers, and few efforts have been made to leverage the hierarchical structure of categories.

In this paper, we introduce hierarchical deep CNNs (HD-CNNs) by embedding deep CNNs into a category hierarchy. An HD-CNN separates easy classes using a coarse category classifier while distinguishing difficult classes using fine category classifiers. During HD-CNN training, component-wise pretraining is followed by global finetuning with a multinomial logistic loss regularized by a coarse category consistency term.

In addition, conditional executions of fine category classifiers and layer parameter compression make HD-CNNs scalable for large-scale visual recognition. We achieve state-of-the-art results on both CIFAR100 and large-scale ImageNet 1000-class benchmark datasets. In our experiments, we build up three different HD-CNNs and they lower the top-1 error of the standard CNNs by 2.65%, 3.1% and 1.1%, respectively.

This paper presents a case study of using distributed word representations, word2vec in particular, for improving performance of Named Entity Recognition for the e-Commerce domain. We also demonstrate that distributed word representations trained on a smaller amount of in-domain data are more effective than word vectors trained on very large amount of out-of-domain data, and that their combination gives the best results.

The performance of automatic speech recognition system degrades significantly when the incoming audio differs from training data. Maximum likelihood linear regression has been widely used for unsupervised adaptation, usually in a multiple-pass recognition process. Here we present a novel adaptation framework for which the offline, supervised, high-quality adaptation is applied to clustered channel/speaker conditions that are defined with automatic and manual clustering of the training data. Upon online recognition, each speech segment is classified into one of the training clusters in an unsupervised way, and the corresponding top acoustic models are used for recognition. Recognition lattice outputs are combined. Experiments are performed on the Wall Street Journal data, and a 37.5% relative reduction of Word Error Rate is reported. The proposed approach is also compared with a general speaker adaptive training approach.