Learning compact class codes for fast inference in large multi class classification

摘要: We describe a new approach for classification with very large number of classes where we assume some class similarity information is available, e.g. through hierarchical organization. The proposed method learns compact binary code using such an existing defined on classes. Binary classifiers are then trained this and decoding performed simple nearest neighbor rule. This strategy, related to Error Correcting Output Codes methods, shown perform similarly or better than the standard efficient one-vs-all approach, much lower inference complexity.