Such an approach is sometimes named webly supervised learning (Chen, Gupta, 2015, Joulin, van der Maaten, Jabri, Vasilache, 2016, Zhang, Shang, Luan, Yang, Chua, 2015).Ĭonvnets became the de facto representation learning method in image classification thanks to its excellent generalization ability. This paper proposes an alternative to train deep neural networks using massive amount of unannotated Web images. For the most part, those neural network models are supervised, which require lots of labelled training data hence pose scalability challenges. Recently the resurgence of neural networks (Bengio, Lamblin, Popovici, Larochelle, et al., 2007, Hinton, 2005, Hinton, Osindero, Teh, 2006) has first led to a revolution in computer vision (Ciresan, Meier, Masci, Schmidhuber, 2012, Krizhevsky, Sutskever, Hinton, 2012, Razavian, Azizpour, Sullivan, Carlsson, 2014, Simonyan, Zisserman, Szegedy, Liu, Jia, Sermanet, Reed, Anguelov, Erhan, Vanhoucke, Rabinovich, 2015), as well as in other areas including reinforcement learning (Mnih et al., 2013), speech recognition (Graves et al., 2013), and natural language processing (Mikolov et al., 2013). The effectiveness of our approach is shown by the good generalization of the learned representations with new six public datasets.įor a long time the vision community has been striving for the quest of creating human-like intelligent systems.
Our experiments are conducted at several data scales, with different choices of network architecture, and alternating between different data preprocessing techniques. We instead train convolutional networks in a supervised setting but use weakly labelled data which are large amounts of unannotated Web images downloaded from Flickr and Bing. Nonetheless they are less performant than supervised methods partly because the loss function used in unsupervised methods, for instance Euclidean loss, failed to guide the network to learn discriminative features and ignore unnecessary details. There have been efforts to train neural networks such as autoencoders with respect to either unsupervised or semi-supervised settings.
The fact that deep networks are hungry for labelled data limits themselves from extracting valuable information of Web images which are abundant and cheap.
The keep-growing content of Web images is probably the next important data source to scale up deep neural networks which recently surpass human in image classification tasks.
0 kommentar(er)
