Bridging the Gap Between Semantic and User Preference Spaces for Multi-modal Music Representation Learning

Recent works of music representation learning mainly focus on learning acoustic music representations with unlabeled audios or further attempt to acquire multi-modal music representations with scarce annotated audio-text pairs. They either ignore the language semantics or rely on labeled audio datasets that are difficult and expensive to create. Moreover, merely modeling semantic space usually fails to achieve satisfactory performance on music recommendation tasks since the user preference space is ignored. In this paper, we propose a novel Hierarchical Two-stage Contrastive Learning (HTCL) method that models similarity from the semantic perspective to the user perspective hierarchically to learn a comprehensive music representation bridging the gap between semantic and user preference spaces. We devise a scalable audio encoder and leverage a pre-trained BERT model as the text encoder to learn audio-text semantics via large-scale contrastive pre-training. Further, we explore a simple yet effective way to exploit interaction data from our online music platform to adapt the semantic space to user preference space via contrastive fine-tuning, which differs from previous works that follow the idea of collaborative filtering. As a result, we obtain a powerful audio encoder that not only distills language semantics from the text encoder but also models similarity in user preference space with the integrity of semantic space preserved. Experimental results on both music semantic and recommendation tasks confirm the effectiveness of our method.

Person re-identification via efficient inference in fully connected CRF

In this paper, we address the problem of person re-identification problem, i.e., retrieving instances from gallery which are generated by the same person as the given probe image. This is very challenging because the person's appearance usually undergoes significant variations due to changes in illumination, camera angle and view, background clutter, and occlusion over the camera network. In this paper, we assume that the matched gallery images should not only be similar to the probe, but also be similar to each other, under suitable metric. We express this assumption with a fully connected CRF model in which each node corresponds to a gallery and every pair of nodes are connected by an edge. A label variable is associated with each node to indicate whether the corresponding image is from target person. We define unary potential for each node using existing feature calculation and matching techniques, which reflect the similarity between probe and gallery image, and define pairwise potential for each edge in terms of a weighed combination of Gaussian kernels, which encode appearance similarity between pair of gallery images. The specific form of pairwise potential allows us to exploit an efficient inference algorithm to calculate the marginal distribution of each label variable for this dense connected CRF. We show the superiority of our method by applying it to public datasets and comparing with the state of the art.