Junhai Xu

Audio anti-spoofing algorithms are widely deployed to defend against spoofing attacks, yet they often fail to detect unseen attacks. Although unsupervised domain adaptation (UDA) offers the potential to address this challenge, existing methods struggle with the large intra-class variability and complex distribution structures in target domains caused by the diversity of speech and attack types. In contrast, optimal transport (OT) leverages the geometric structure of intra-class distributions to measure discrepancies between probability distributions. The effectiveness of OT relies on the discriminability of data within target domains. However, in real-world scenarios involving multiple target domains, these domains often overlap in feature space, leading to the negative transport problem in OT. To overcome these domain mismatches in anti-spoofing, we propose the Multi-Sinkhorn Teacher Knowledge Aggregation (MSTKA) framework. Initially, to avoid interference between target domains during alignment, we use OT to adapt the source model to each target domain independently, thereby reducing negative transport. This adaptation involves constructing an OT cost matrix based on sentence-level representations of cross-domain samples and training an expert model for each target domain. Subsequently, we aggregate the knowledge from these expert models into a unified student model, enabling it to generalize across multiple target domains. Since spoofing cues could be distributed across different temporal scales, we align the student model's representations at multiple time scales with the teacher model's sentence-level representations to enhance the effectiveness of knowledge distillation. Multi-target adaptation experiments on eleven data sets demonstrate that our framework achieves state-of-the-art performance in audio anti-spoofing.

Video-based remote physiological measurement, enabling non-contact monitoring of physiological signals through facial videos, demonstrates promising prospects in remote medicine and health monitoring. To take advantage of long-term dependency information from cardiac activity, previous methods commonly rely on long video sequences for measurement. These approaches are highly susceptible to various internal and external factors and struggle to capture fluctuations in physiological signals over extended measurement periods reliably. Unlike previous long-form video-based methods, we propose a novel adaptive knowledge distillation framework that utilizes short-form videos. This approach aims to mitigate the impact of internal and external factors while accurately capturing physiological signal fluctuations by reducing the required video length for precise measurement. Specifically, we first propose a long-short-term distillation framework that uses a teacher model to capture long-term dependencies from long-form videos and leverages this capability to guide the training of the student model on short-form videos. Subsequently, we propose a frame-level dynamic prior mechanism (FDPM) that replaces student features similar to teacher features with the corresponding teacher features and dynamically adjusts their proportions. Furthermore, we design an instance adaptive distillation mechanism (IADM) module that dynamically regulates the distillation temperature and the loss weight to develop an optimal distillation strategy for each sample. Finally, we conduct systematic experiments across three benchmark datasets and various video lengths. Experimental results demonstrate the effectiveness of our method, especially in short-form video scenarios.

Deep learning frameworks utilizing convolutional neural networks (CNNs) have frequently been used for brain age prediction and have achieved outstanding performance. Nevertheless, deep learning remains a black box as it is hard to interpret which brain parts contribute significantly to the predictions. To tackle this challenge, we first trained a lightweight, fully CNN model for brain age estimation on a large sample data set (N = 3054, age range = [8,80 years]) and tested it on an independent data set (N = 555, age range = [8,80 years]). We then developed an interpretable scheme combining network occlusion sensitivity analysis (NOSA) with a fine‐grained human brain atlas to uncover the learned invariance of the model. Our findings show that the dorsolateral, dorsomedial frontal cortex, anterior cingulate cortex, and thalamus had the highest contributions to age prediction across the lifespan. More interestingly, we observed that different regions showed divergent patterns in their predictions for specific age groups and that the bilateral hemispheres contributed differently to the predictions. Regions in the frontal lobe were essential predictors in both the developmental and aging stages, with the thalamus remaining relatively stable and saliently correlated with other regional changes throughout the lifespan. The lateral and medial temporal brain regions gradually became involved during the aging phase. At the network level, the frontoparietal and the default mode networks show an inverted U‐shape contribution from the developmental to the aging stages. The framework could identify regional contributions to the brain age prediction model, which could help increase the model interpretability when serving as an aging biomarker.

Audio anti-spoofing algorithms struggle with fake samples from unseen spoofing techniques, even when trained with diverse data sets or data augmentation strategies. Unsupervised domain adaptation (UDA) algorithms have the potential to mitigate this challenge. Typically, UDA assumes that the source and target domains are distinct distributions with clear boundaries and seeks to align model representations between them. However, in anti-spoofing, various spoofing algorithms could cause the distributions of the generated samples to overlap, resulting in unclear domain boundaries. This hinders UDA algorithms from effectively measuring and aligning domain discrepancies. Moreover, forcibly aligning samples with significant discrepancies could diminish the model’s discriminative capability. To solve this problem, we propose a domain attention algorithm with optimal transport (OT), termed Sinkhorn Domain Attention (SHDA). Unlike traditional attention mechanisms, SHDA identifies the optimal transfer plan by analyzing the global probability differences among cross-domain samples. Specifically, we first extract audio representations from various domains to compute the overall cost matrix between the source and target domains. Next, we employ Sinkhorn’s iteration to calculate the OT coupling matrix, where cross-domain samples with minor differences receive higher transfer weights, while those with substantial differences receive lower weights. Finally, we use the coupling and cost matrices to compute the adaptation loss, effectively transferring the anti-spoofing model from multiple sources to the target domain. We conducted eight cross-domain experiments using eleven well-known anti-spoofing corpora. The results indicate that our label-free SHDA surpassed the state-of-the-art model by 40%.

In recent years, numerous studies have demonstrated a close connection between human diseases and the regulation of non-coding RNAs (ncRNAs). Predicting potential ncRNAs associated with disease can help provide critical information for diagnosis and treatment of disease, leading to better disease analysis and prevention. Building good algorithms for predicting associations between ncRNAs and disease is critical. Many current algorithms have poor performance in identifying the association between ncRNAs and diseases. As a method for predicting the association between ncRNAs and diseases, we develop a Matrix Factorization method based on the Correntropy Induced Loss (C-loss) function (C-lossMF). In our model, we first construct ncRNA similarity matrix and disease similarity matrix by considering some important similarity information, and extract effective information of ncRNA and disease from them. Next, we perform matrix decomposition of ncRNA-disease association matrix and apply $L2$ loss and C-loss. Then we add collaborative regularization of RNA similarity matrix and the collaborative regularization of disease similarity matrix to take full advantage of the information in the similarity matrix. In particular, we propose a method that combines semi-quadratic optimization and gradient descent to optimize the model. In the experiments, we utilize the five-fold cross validation method on four datasets to evaluate the performance of C-lossMF. Comparing this model with other advanced models, the results show that it performs better.

Speech emotion recognition (SER) has widespread applications in human-computer interaction. However, the performance of SER models often drops in domain mismatch scenarios. Although existing unsupervised domain adaptation (UDA) algorithms could mitigate the domain mismatch problem by aligning feature distributions between domains, they do not consider the distribution relationship between emotion categories, thereby decreasing the capability to discriminate emotions. Moreover, directly aligning the probability distributions of different domains overlooks the emotional structure information contained in the target domain samples. To overcome this limitation, this paper proposes a novel UDA for cross-domain SER, termed Optimal Transport (OT) with Class Structure Exploration (OTCSE). OTCSE aims to measure the global probability distribution distance (GPDD) differences while considering the distribution relationship between emotional categories. Additionally, it explores the intrinsic structure information in the target samples. Specifically, we first embed joint OT into a deep SER framework to measure the GPDD between domains. Second, we propose to use a self-supervised learning (SSL)-based domain exploration module in the OT adaptation process to assist in exploring class structure information. Finally, we propose a two-step optimization strategy that allows OTCSE to update the model parameters of the SER and SSL modules end-to-end while solving the optimal transport coupling based on Sinkhorn's iteration algorithm. Experimental results in cross-domain SER demonstrate that the proposed OTCSE outperformed state-of-the-art UDAs. Further, we explore the complementarity of GPDD alignment and class structure exploration. This synergy alleviated the negative transport problem in OT and enhanced the efficiency of representation learning in SSL.

Cross-domain speaker recognition (SR) can be improved by unsupervised domain adaptation (UDA) algorithms. UDA algorithms often reduce domain mismatch at the cost of decreasing the discrimination of speaker features. In contrast, optimal transport (OT) has the potential to achieve domain alignment while preserving the speaker discrimination capability in UDA applications; however, naively applying OT to measure global probability distribution discrepancies between the source and target domains may induce negative transports where samples belonging to different speakers are coupled in transportation. These negative transports reduce the SR model's discriminative power, degrading the SR performance. This paper proposes a coupling-regularized optimal transport (CROT) algorithm for cross-domain SR to reduce the negative transport during UDA. In the proposed CROT, two consecutive processing modules regularize the coupling paths for the OT solution: a progressive inter-speaker constraint (PISC) module and a coupling-smoothed regularization (CSR) module. The PISC, designed as a pseudo-label memory bank with curriculum learning, is first applied to select valid samples to guarantee that coupling samples are from the same speaker. The CSR, designed to control the information entropy of the coupling paths further, reduces the effect of negative transport in UDA. To evaluate the effectiveness of the proposed algorithm, cross-domain SR experiments were conducted under different target domains, speaker encoders, corpora, and acoustic features. Experimental results showed that CROT achieved a 50% relative reduction in equal error rates compared to conventional OT-based UDAs, outperforming the state-of-the-art UDAs.

Medical image segmentation is a critical task in the healthcare field. While deep learning techniques have shown promise in this area, they often require a large number of accurately labeled images. To address this issue, semisupervised learning has emerged as a potential solution by reducing the reliance on precise annotations. Among these approaches, the student-teacher framework has garnered attention, but it is limited in its reliance solely on the teacher model for information. To overcome this limitation, we propose a prototype-based mutual consistency learning (PMCL) framework. This framework utilizes two branches that learn from each other, incorporating supervision loss and consistency loss to adapt to minor data perturbations and structural differences. By employing prototype consistency learning, we are able to achieve reliable consistency loss. Our experiments on three public medical image datasets demonstrate that PMCL outperforms other state-of-the-art methods, indicating its potential in semisupervised medical image segmentation. Our framework has the potential to assist medical professionals in enhancing their diagnoses and delivering improved patient care.