Towards a universal continuous knowledge base

In artificial intelligence (AI), knowledge is the information required by an intelligent system to accomplish tasks. While traditional bases use discrete, symbolic representations, detecting encoded in continuous representations learned from data has received increasing attention recently. this work, we propose a method for building base (CKB) that can store imported multiple, diverse neural networks. The key idea of our approach define interface each network and cast transferring as function simulation problem. Experiments on text classification show promising results: CKB imports single model then exports new model, achieving comparable performance with original model. More interesting, import multiple models base, which fused exported back higher accuracy than With CKB, it also easy achieve distillation transfer learning. Our work opens door universal collect, store, organize all various networks trained different AI