Quaternion anti-transfer learning for speech emotion recognition

This study explores the benefits of anti-transfer learning with quaternion neural networks for robust, effective, and efficient speech emotion recognition. Anti-transfer learning selectively promotes task invariance through the introduction of a deep feature loss at training time. It has been shown to improve the performance of speech emotion recognition models by encouraging the independence of emotion predictions from specific uttered words and characteristics of the speaker's voice. However, the improved accuracy comes at a cost of increased computation time and memory requirements. In order to reduce the resource demand of anti-transfer, we propose to exploit quaternion-valued processing. We design, implement, and evaluate the use of quaternion anti-transfer learning on the basis of the VGG16 architecture and quaternion embeddings on multiple datasets for different speech emotion recognition task setups. The effectiveness of this approach depends on the layer where it is applied, with early layers offering a good compromise between performance gain and resource requirements. Our results show that anti-transfer in the quaternion domain can enhance generalisation while reducing the model's demand for computation and memory.