Efficient Learning Framework for Training Deep Learning Models with Limited SupervisionGhasedi Dizaji, Kamran (2021) Efficient Learning Framework for Training Deep Learning Models with Limited Supervision. Doctoral Dissertation, University of Pittsburgh. (Unpublished)
AbstractIn recent years, deep learning has shown tremendous success in different applications, however these modes mostly need a large labeled dataset for training their parameters. In this work, we aim to explore the potentials of efficient learning frameworks for training deep models on different problems in the case of limited supervision or noisy labels. For the image clustering problem, we introduce a new deep convolutional autoencoder with an unsupervised learning framework. We employ a relative entropy minimization as the clustering objective regularized by the frequency of cluster assignments and a reconstruction loss. In the case of noisy labels obtained by crowdsourcing platforms, we proposed a novel deep hybrid model for sentiment analysis of text data like tweets based on noisy crowd labels. The proposed model consists of a crowdsourcing aggregation model and a deep text autoencoder. We combine these sub-models based on a probabilistic framework rather than a heuristic way, and derive an efficient optimization algorithm to jointly solve the corresponding problem. In order to improve the performance of unsupervised deep hash functions on image similarity search in big datasets, we adopt generative adversarial networks to propose a new deep image retrieval model, where the adversarial loss is employed as a data-dependent regularization in our objective function. We also introduce a balanced self-paced learning algorithm for training a GAN-based model for image clustering, where the input samples are gradually included into training from easy to difficult, while the diversity of selected samples from all clusters are also considered. In addition, we explore adopting discriminative approaches for unsupervised visual representation learning rather than the generative algorithms, such as maximizing the mutual information between an input image and its representation and a contrastive loss for decreasing the distance between the representations of original and augmented image data. Share
Details
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|
![[feed]](/images/feed-icon-32x32.png){kind=icon}