The disclosed technique provides a pseudo-recurrent neural network (QRNN) that alternates between convolutional layers that apply in parallel over time steps and minimalist recursive pooling layers that apply in parallel across feature dimensions.
展开▼
