| Popis: |
Modern deep neural networks perform well on a variety of tasks where a large amount of training data is available, but in low data settings, humans still outperform neural networks significantly. An important reason for this is the ability to reuse knowledge from other problems. Guzdial and Riedl proposed the approach of conceptual expansion[1], a general and efficient technique for reusing the knowledge of trained models for new tasks. In this paper we propose modifications to the original method of conceptual expansion to greatly increase the accuracy of the approach. |
