GlanceNets - efficient convolutional neural networks with adaptive hard example mining

Despite the success of CNNs, it is impeded to deploy such deep CNN models in real-time tasks due to high computational complexity. To address the problem, we propose GlanceNets with several bypasses (Figure 1). In modern CNNs, it is believed that shallow layers provide lower-level features, whereas deep layers correspond to higherlevel features. However, it is not always necessary to classify a sample with the highest-level feature. In many cases, easy samples can be correctly classified with low-level features, just as one can recognize common items at a glance. Such observation is the key motivation of proposed GlanceNets in this study.

Recommended citation: Sun Hanqing, Pang Yanwei. GlanceNets - efficient convolutional neural networks with adaptive hard example mining. Science China Information Sciences, 2018, 61(10): 109101.
Download Paper