Abstract: Current works on multimodal facial expression recognition typically require paired visible and thermal facial images. Although visible cameras are readily available in our daily life, thermal cameras are expensive and less prevalent. It is costly to collect a large quantity of synchronous visible and thermal facial images. To tackle this paired training data bottleneck, we propose an unpaired multimodal facial expression recognition method, which makes full use of the massive number of unpaired visible and thermal images by utilizing thermal images to construct better image representations and classifiers for visible images during training. Specifically, two deep neural networks are trained from visible and thermal images to learn image representations and expression classifiers for two modalities. Then, an adversarial strategy is adopted to force statistical similarity between the learned visible and thermal representations, and to minimize the distribution mismatch between the predictions of the visible and thermal images. Through adversarial learning, the proposed method leverages thermal images to construct better image representations and classifiers for visible images during training, without the requirement of paired data. A decoder network is built upon the visible hidden features in order to preserve some inherent features of the visible view. We also take the variability of the different images�_transferability into account via adaptive classification loss. During testing, only visible images are required and the visible network is used. Thus, the proposed method is appropriate for real-world scenarios, since thermal imaging is rare in these instances. Experiments on two benchmark multimodal expression databases and three visible facial expression databases demonstrate the superiority of the proposed method compared to state-of-the-art methods.