RESUMO
Low-rankness plays an important role in traditional machine learning but is not so popular in deep learning. Most previous low-rank network compression methods compress networks by approximating pretrained models and retraining. However, the optimal solution in the Euclidean space may be quite different from the one with low-rank constraint. A well-pretrained model is not a good initialization for the model with low-rank constraints. Thus, the performance of a low-rank compressed network degrades significantly. Compared with other network compression methods such as pruning, low-rank methods attract less attention in recent years. In this article, we devise a new training method, low-rank projection with energy transfer (LRPET), that trains low-rank compressed networks from scratch and achieves competitive performance. We propose to alternately perform stochastic gradient descent training and projection of each weight matrix onto the corresponding low-rank manifold. Compared to retraining on the compact model, this enables full utilization of model capacity since solution space is relaxed back to Euclidean space after projection. The matrix energy (the sum of squares of singular values) reduction caused by projection is compensated by energy transfer. We uniformly transfer the energy of the pruned singular values to the remaining ones. We theoretically show that energy transfer eases the trend of gradient vanishing caused by projection. In modern networks, a batch normalization (BN) layer can be merged into the previous convolution layer for inference, thereby influencing the optimal low-rank approximation (LRA) of the previous layer. We propose BN rectification to cut off its effect on the optimal LRA, which further improves the performance. Comprehensive experiments on CIFAR-10 and ImageNet have justified that our method is superior to other low-rank compression methods and also outperforms recent state-of-the-art pruning methods. For object detection and semantic segmentation, our method still achieves good compression results. In addition, we combine LRPET with quantization and hashing methods and achieve even better compression than the original single method. We further apply it in Transformer-based models to demonstrate its transferability. Our code is available at https://github.com/BZQLin/LRPET.
RESUMO
OBJECTIVES: The study aimed to develop and validate machine learning models for case difficulty prediction in endodontic microsurgery, assisting clinicians in preoperative analysis. METHODS: The cone-beam computed tomographic images were collected from 261 patients with 341 teeth and used for radiographic examination and measurement. Through linear regression (LR), support vector regression (SVR), and extreme gradient boosting (XGBoost) algorithms, four models were established according to different loss functions, including the L1-loss LR model, L2-loss LR model, SVR model and XGBoost model. Five-fold cross-validation was applied in model training and validation. Explained variance score (EVS), coefficient of determination (R2), mean absolute error (MAE), mean squared error (MSE) and median absolute error (MedAE) were calculated to evaluate the prediction performance. RESULTS: The MAE, MSE and MedAE values ââof the XGBoost model were the lowest, which were 0.1010, 0.0391 and 0.0235, respectively. The EVS and R2 values ââof the XGBoost model were the highest, which were 0.7885 and 0.7967, respectively. The factors used to predict the case difficulty in endodontic microsurgery were ordered according to their relative importance, including lesion size, the distance between apex and adjacent important anatomical structures, root filling density, root apex diameter, root resorption, tooth type, tooth length, root filling length, root canal curvature and the number of root canals. CONCLUSIONS: The XGBoost model outperformed the LR and SVR models on all evaluation metrics, which can assist clinicians in preoperative analysis. The relative feature importance provides a reference to develop the scoring system for case difficulty assessment in endodontic microsurgery. CLINICAL SIGNIFICANCE: Preoperative case assessment is a crucial step to identify potential risks and make referral decisions. Machine learning models for case difficulty prediction in endodontic microsurgery can assist clinicians in preoperative analysis efficiently and accurately.
Assuntos
Microcirurgia , Tratamento do Canal Radicular , Humanos , Microcirurgia/métodos , Tratamento do Canal Radicular/métodos , Tomografia Computadorizada de Feixe Cônico , AlgoritmosRESUMO
Electrodermal activity (EDA) sensor is emerging non-invasive equipment in affect detection research, which is used to measure electrical activities of the skin. Knowledge graphs are an effective way to learn representation from data. However, few studies analyzed the effect of knowledge-related graph features with physiological signals when subjects are in non-similar mental states. In this paper, we propose a model using deep learning techniques to classify the emotional responses of individuals acquired from physiological datasets. We aim to improve the execution of emotion recognition based on EDA signals. The proposed framework is based on observed gender and age information as embedding feature vectors. We also extract time and frequency EDA features in line with cognitive studies. We then introduce a sophisticated weighted feature fusion method that combines knowledge embedding feature vectors and statistical feature (SF) vectors for emotional state classification. We finally utilize deep neural networks to optimize our approach. Results obtained indicated that the correct combination of Gender-Age Relation Graph (GARG) and SF vectors improve the performance of the valence-arousal emotion recognition system by 4 and 5% on PAFEW and 3 and 2% on DEAP datasets.
RESUMO
Emotion recognition from affective brain-computer interfaces (aBCI) has garnered a lot of attention in human-computer interactions. Electroencephalographic (EEG) signals collected and stored in one database have been mostly used due to their ability to detect brain activities in real time and their reliability. Nevertheless, large EEG individual differences occur amongst subjects making it impossible for models to share information across. New labeled data is collected and trained separately for new subjects which costs a lot of time. Also, during EEG data collection across databases, different stimulation is introduced to subjects. Audio-visual stimulation (AVS) is commonly used in studying the emotional responses of subjects. In this article, we propose a brain region aware domain adaptation (BRADA) algorithm to treat features from auditory and visual brain regions differently, which effectively tackle subject-to-subject variations and mitigate distribution mismatch across databases. BRADA is a new framework that works with the existing transfer learning method. We apply BRADA to both cross-subject and cross-database settings. The experimental results indicate that our proposed transfer learning method can improve valence-arousal emotion recognition tasks.
RESUMO
GoDec is an efficient low-rank matrix decomposition algorithm. However, optimal performance depends on sparse errors and Gaussian noise. This paper aims to address the problem that a matrix is composed of a low-rank component and unknown corruptions. We introduce a robust local similarity measure called correntropy to describe the corruptions and, in doing so, obtain a more robust and faster low-rank decomposition algorithm: GoDec+. Based on half-quadratic optimization and greedy bilateral paradigm, we deliver a solution to the maximum correntropy criterion (MCC)-based low-rank decomposition problem. Experimental results show that GoDec+ is efficient and robust to different corruptions including Gaussian noise, Laplacian noise, salt & pepper noise, and occlusion on both synthetic and real vision data. We further apply GoDec+ to more general applications including classification and subspace clustering. For classification, we construct an ensemble subspace from the low-rank GoDec+ matrix and introduce an MCC-based classifier. For subspace clustering, we utilize GoDec+ values low-rank matrix for MCC-based self-expression and combine it with spectral clustering. Face recognition, motion segmentation, and face clustering experiments show that the proposed methods are effective and robust. In particular, we achieve the state-of-the-art performance on the Hopkins 155 data set and the first 10 subjects of extended Yale B for subspace clustering.