2014/11/17 by Sotirios Chatzis, Chatzis, Sotirios P. · 1 citation
Computer Science · Engineering · #Domain Adaptation and Few-Shot Learning #Industrial Vision Systems and Defect Detection #Blind Source Separation Techniques
paper · pdf · doi:10.48550/arxiv.1411.4423
Unsupervised feature learning algorithms based on convolutional formulations\nof independent components analysis (ICA) have been demonstrated to yield\nstate-of-the-art results in several action recognition benchmarks. However,\nexisting approaches do not allow for the number of latent components (features)\nto be automatically inferred from the data in an unsupervised manner. This is a\nsignificant disadvantage of the state-of-the-art, as it results in considerable\nburden imposed on researchers and practitioners, who must resort to tedious\ncross-validation procedures to obtain the optimal number of latent features. To\nresolve these issues, in this paper we introduce a convolutional nonparametric\nBayesian sparse ICA architecture for overcomplete feature learning from\nhigh-dimensional data. Our method utilizes an Indian buffet process prior to\nfacilitate inference of the appropriate number of latent features under a\nhybrid variational inference algorithm, scalable to massive datasets. As we\nshow, our model can be naturally used to obtain deep unsupervised hierarchical\nfeature extractors, by greedily stacking successive model layers, similar to\nexisting approaches. In addition, inference for this model is completely\nheuristics-free; thus, it obviates the need of tedious parameter tuning, which\nis a major challenge most deep learning approaches are faced with. We evaluate\nour method on several action recognition benchmarks, and exhibit its advantages\nover the state-of-the-art.\n