Knowledge distillation, as a model compression technique , has been widely applied in artificial intelligence to improve the efficiency of deep learning models, especially in resource-constrained environments. Considering that logit contains more decision information compared to the intermediate feature maps, fully logits guided distillation is proposed, which allows student networks to have better access to the guidance from both the intermediate and decision levels of the teacher’s network. Intermediate feature logicalization is designed, which perform logit transformations on the intermediate feature maps to obtain intermediate decision information. A logit matrisation strategy is proposed, which aim to capture inter-class information of the logits. Furthermore, cross layer distillation is presented in order to enable the final logit of the teacher to provide guidance to the intermediate layers of the student. The proposed mechanism can be embedded into State-of-the-art distillation frameworks to further improve the accuracy. Experiments conducted on the CIFAR-10, CIFAR-100, and ImageNet datasets demonstrate the effectiveness of the proposed method. Image classification accuracy was used as the evaluation metric, and the results show that the proposed method improves accuracy by an average of 1.84%, with a best improvement of 3.19% over the baseline models. Code is available at https://github.com/YiqinWang-JN/FLGD .

Identification of cancer subtypes plays an important role in revealing useful insights into disease pathogenesis and advancing personalized therapy. This endeavor significantly relies on multi-omics data, whose pivotal role in subtype classification has been widely recognized and applied in current biomedical research and clinical practice. Intricate high-dimensional data in each omics contain significant number of discriminative features as well as noise information. Meanwhile, the involvement of diverse omics in classification tasks varies significantly, contributing to varying degrees of importance for the classification tasks. To take advantage of the unique discriminative information embedded in multi omics data, it is necessary to integrate multi-omics data into a feature space that emphasizes discriminative information maximally, while effectively disregarding irrelevant information. In this work, we propose Collaborative Attention Contrast Learning (CACL) framework, which integrates a genetic attention module (GAM) to capture key intra-omics features and an omics attention module (OAM) to enhance inter-omics relationships and optimizes the collaborative attention models through the strategic utilization of a contrastive loss function. This optimization strategy empowers the algorithm to extract multi-omics fusion features with enhanced discriminative ability, ultimately leading to a remarkable improvement in clustering performance. Experiments conducted on several representative multi-omics cancer datasets have demonstrated that our proposed method outperforms a number of state-of-the-art methods. Furthermore, the findings indicate that our method is capable of identifying clinically significant subgroups across diverse cancer types.

A multi‐directional salience based similarity evaluation for person re‐identification (re‐id) is presented. After distribution analysis for salience consistency between image pairs, a similarity between matched patches is established by weighted fusion of multi‐directional salience. The weight of saliency in each direction is obtained using metric learning by means of structural support vector machines ranking. The discriminative and accurate performance of re‐id is achieved. Compared with existing salience based person matching framework, the proposed method achieves higher re‐id rate with multi‐directional salience based similarity evaluation.

Expression face neutralisation helps to improve the performance of expressive face recognition with one single neutral sample in gallery per subject. For learning‐based expression neutralisation, the virtual neutral face totally relies on training samples, which removes person‐specific characters from the neutralised face. Bilinear kernel rank reduced regression (BKRRR) algorithm is designed in a virtual subspace to simultaneously and efficiently generate both virtual expressive and neutral images from training samples. An expression mask is then established using grey and gradient differences of the two images. The test expression image is transformed to neutral template by piece‐wise affine warp (PAW). Using the virtual BKRRR neutral image as source, the PAW image as destination and the area covered by expression mask as clone area, an image fusion strategy based on Poisson equation is then designed, which achieves virtual neutralised face image with person‐specific characters preserved. From experiments on the CMU Multi‐PIE databases, it could be observed that the neutral faces synthesised by the proposed method could effectively approximate the real ground truth expressive faces, and greatly improve the performance of classic face recognition algorithms on expression variant problems.

A hybrid active appearance model (AAM) called sequential AAM (SAAM) based on online instance learning is presented. The subspace of the subject-specific AAM component is initially learned with sequential registration results of first frames, and is periodically updated through incremental principal component analysis and online instance fitting process. A drift correction component of the AAM is also updated during tracking by selecting previous ‘good fitting’ frame as a reference image. With the model, facial features can be tracked in a video given theirs locations in the first frame and no other information. Experiments show improved fitting accuracy and computation cost compared with other state-of-the-art AAM.