Acceleration of Predictive Elastic Weight Consolidation Using Node Pruning

초록

Significant performance improvements have been mode through various studies on deep learning. When a network is deeper, it makes faster convergence to the local minimum. Therefore, a large number of studies have been conducted on deep networks to design networks with high performance. However, as a net-work becomes deeper, the network model also occupies more storage, resulting in difficulties for mobile environment applications. Thus, this study accelerated the Predictive Elastic Weight Consolidation without performance degradation through node pruning using Fisher Information. The experimental results showed that learning speed improved without causing performance degradation.