Guided Dropout: Improving Deep Networks Without Increased Computation

Deep convolution neural networks are going deeper and deeper. However, the complexity of models is prone to overfitting in training. Dropout, one crucial tricks, prevents units from co-adapting too much by randomly dropping neurons during It effectively improves performance deep but ignores importance differences between neurons. To optimize this issue, paper presents a new dropout method called guided dropout, which selects switch off according kernel preserves informative uses an unsupervised clustering algorithm cluster similar each hidden layer, certain probability within cluster. Thereby would preserve layer with different roles while maintaining model’s scarcity generalization, role learning features. We evaluated our approach compared two standard on three well-established public object detection datasets. Experimental results multiple datasets show that proposed has been improved false positives, precision-recall curve average precision without increasing amount computation. can be seen increased thanks shallow networks. The concept beneficial other vision tasks.