Academia vicarious in ICLR 2017 submission of unsupervised learning papers, the proposed combination huangshexiaoshuo

The academic | Vicarious submitted in ICLR 2017 unsupervised learning, proposed learning hierarchy combination characteristics of Sohu technology selected from Open Review machine of the heart: Li Yazhou Wu Pan, compiled in the field of artificial intelligence star startups Vicarious has been the subject of great concern in the industry, Amazon CEO Bezos Zuckerberg, Facebook CEO, Salesforce CEO and Marc Benioff Box CEO Aaron Levie are the investors. Recently, Vicarious published a new paper on unsupervised learning, which has been submitted to ICLR 2017. Click to read the original text to download this paper. Abstract we introduce a hierarchical combination network (hierarchical compositional network, HCN), it is a generative model oriented (directed generative model), in the case of unsupervised discovery and solve the two value image (binary images) set construction module. These building blocks are defined as some features of hierarchical network after layer (in a special form of arrangement) two valued feature combination. From a high level, HCN is similar to a pool of S belief networks (sigmoid belief network). The inference and learning in HCN is very challenging, and the existing variational approximation method is not very satisfactory. A major contribution of the study was found: the use of special arrangements (no EM) max-product message (MPMP), the two issues mentioned earlier can be solved. Moreover, the use of MPMP as a new HCN inference engine makes the task more simple: adding supervision information and classification of image or perform image restoration — all correspond to a known model of some variables is fixed on the value of the model, and then run MPMP in the remaining part. When used for classification, the fast inference of HCN is almost the same as that of convolution neural network with linear activation function and two valued weight. However, the characteristics of HCN are very different in quality. Key words: unsupervised learning © this paper compiled by the heart of the machine, reproduced please contact the public number to obtain authorization. ————————————————? Join the heart of machines (full-time reporter Intern): hr@almosthuman submission or seek reports: editor@almosthuman advertising business cooperation: bd@almosthuman &相关的主题文章:

« »

Comments closed.