Name: Lei Zhang
Affiliation: Guangdong University of Petrochemical Technology.
Invited Talk: Compact Network with Cosine Activation
Abstract: A new learning architecture named compact network with cosine activation (CNCA) is proposed. CNCA is derived from kernel approximation and establishes a nonlinear hidden layer with the cosine activation function. By carefully selecting normalization strategy and combined with low rank constraint, CNCA exhibits excellent performance in face alignment task. Extensive experiments on five in-the-wild face alignment datasets show that CNCA delivers high performance and consistently exceeds state-of-the-art methods. By seamlessly connecting with convolutional neural networks (CNNs), and further combined with spatial pyramid matching to fuse various information into one holistic picture, CNCA can also deal with scene classification task. Experiments on the MIT indoor and SUN397 datasets show that CNCA delivers high performance and demonstrates its great effectiveness for scene-classification tasks.