Global Perception Feedback Convolutional Neural Networks

Top-down feedback mechanism is an important module of visual attention for weakly supervised learning. Previous top-down feedback convolutional neural networks often perform local perception during feedback. Inspired by the fact that the visual system is sensitive to global topological properties [1], we propose a global perception feedback convolutional neural network that considers the global structure of visual response during feedback inference. The global perception eliminates “Visual illusions” that are produced in the process of visual attention. It is achieved by simply imposing the trace norm on hidden neuron activations. Particularly, when updating the status of hidden neuron activations during gradient backpropagation, we get rid of some minor constituent in the SVD decomposition, which both ensures the global low-rank structure of feedback information and the elimination of local noise. Experimental results on the ImageNet dataset corroborate our claims and demonstrate the effectiveness of our global perception model.