Information Fusion in Attention Networks Using Adaptive and Multi-level Factorized Bilinear Pooling for Audio-visual Emotion Recognition.