Unsupervised Learning on Neural Network Outputs: With Application in Zero-Shot Learning

2015

39 citations recorded by the source corpus.

The outputs of a trained neural network contain much richer information than just an one-hot classifier. For example, a neural network might give an image of a dog the probability of one in a million of being a cat but it is still much larger than the probability of being a car. To reveal the hidden structure in them, we apply two unsupervised learning algorithms, PCA and ICA, to the outputs of a deep Convolutional Neural Network trained on the ImageNet of 1000 classes. The PCA/ICA embedding of the object classes reveals their visual similarity and the PCA/ICA components can be interpreted as common visual features shared by similar object classes. For an application, we proposed a new zero-shot learning method, in which the visual features learned by PCA/ICA are employed. Our zero-shot learning method achieves the state-of-the-art results on the ImageNet of over 20000 classes.

Citations in the corpus

• Growing Interpretable Part Graphs on ConvNets via Multi-Shot Learning (2016, background): “…on the visualization of CNN units (Zeiler and Fergus 2014; Mahendran and Vedaldi 2015; Simonyan, Vedaldi, and Zisserman 2013; Dosovitskiy and Brox 2015; Aubry and Russell 2015) and analyzed their statistical features (Dosovitskiy et al. 2014; Szegedy et al. 2014; Yosinski et al. 2014; Lu 2015).”
• Visual interpretability for deep learning: a survey (2018, background): “[Lu, 2015; Aubry and Russell, 2015] have computed feature distributions of different categories/attributes in the feature space of a pre-trained CNN.”
• Zero-Shot Recognition via Semantic Embeddings and Knowledge Graphs (2018, background): “Researchers have proposed several approaches on how to use knowledge graphs for object recognition [12, 43, 30, 35, 39, 9, 8, 29, 49, 47, 26].”
• Rethinking Knowledge Graph Propagation for Zero-Shot Learning (2018, unknown): “Previously, prior knowledge has been incorporated in form of semantic descriptions of classes, such as attributes [1, 28, 19] or word embeddings [30, 11, 18], or by using semantic relations such as knowledge graphs [24, 27, 29, 20].”
• Semantic Autoencoder for Zero-Shot Learning (2017, background): “(3) The third group of methods learn an intermediate space where both the feature space and the semantic space are projected to [36, 67, 13].”
• Domain-Invariant Projection Learning for Zero-Shot Recognition (2018, background): “(3) The third group learns an intermediate space as the embedding space, where both the feature space and the semantic space are projected to [31, 68, 8].”