ForestHash: Semantic Hashing with Shallow Random Forests and Tiny Convolutional Networks

摘要

Hash codes are efficient data representations for coping with the evergrowing amounts of data. In this paper, we introduce a random forest semantichashing scheme that embeds tiny convolutional neural networks (CNN) intoshallow random forests, with near-optimal information-theoretic codeaggregation among trees. We start with a simple hashing scheme, where randomtrees in a forest act as hashing functions by setting `1' for the visited treeleaf, and `0' for the rest. We show that traditional random forests fail togenerate hashes that preserve the underlying similarity between the trees,rendering the random forests approach to hashing challenging. To address this,we propose to first randomly group arriving classes at each tree split nodeinto two groups, obtaining a significantly simplified two-class classificationproblem, which can be handled using a light-weight CNN weak learner. Suchrandom class grouping scheme enables code uniqueness by enforcing each class toshare its code with different classes in different trees. A non-conventionallow-rank loss is further adopted for the CNN weak learners to encourage codeconsistency by minimizing intra-class variations and maximizing inter-classdistance for the two random class groups. Finally, we introduce aninformation-theoretic approach for aggregating codes of individual trees into asingle hash code, producing a near-optimal unique hash for each class. Theproposed approach significantly outperforms state-of-the-art hashing methodsfor image retrieval tasks on large-scale public datasets, while performing atthe level of other state-of-the-art image classification techniques whileutilizing a more compact and efficient scalable representation. This workproposes a principled and robust procedure to train and deploy in parallel anensemble of light-weight CNNs, instead of simply going deeper.