Neural Processing Letters最新文献

Knowledge distillation is a model compression technique that transfers knowledge learned by teacher networks to student networks. Existing knowledge distillation methods greatly expand the forms of knowledge, but also make the distillation models complex and symmetric. However, few studies have explored the commonalities among these methods. In this study, we propose a concise distillation framework to unify these methods and a method to construct asymmetric knowledge distillation under the framework. Asymmetric distillation aims to enable differentiated knowledge transfers for different distillation objects. We designed a multi-stage shallow-wide branch bifurcation method to distill different knowledge representations and a grouping ensemble strategy to supervise the network to teach and learn selectively. Consequently, we conducted experiments using image classification benchmarks to verify the proposed method. Experimental results show that our implementation can achieve considerable improvements over existing methods, demonstrating the effectiveness of the method and the potential of the framework.

In this paper, group consensus is investigated for a class of nonlinear multi-agent systems suffered from the DoS attacks. Firstly, a first-order nonlinear multi-agent system is constructed, which is divided into M subsystems and each subsystem has an unique leader. Then a protocol is proposed and a Lyapunov function candidate is chosen. By means of the stability theory, a sufficient criterion, which involves the duration of DoS attacks, coupling strength and control gain, is obtained for achieving group consensus in first-order system. That is, the nodes in each subsystem can track the leader of that group. Furthermore, the result is extended to nonlinear second-order multi-agent systems and the controller is also improved to obtain sufficient conditions for group consensus. Additionally, the lower bounds of the coupling strength and average interval of DoS attacks can be determined from the obtained sufficient conditions. Finally, several numerical simulations are presented to explain the effectiveness of the proposed controllers and the derived theoretical results.

Fuzzy cognitive maps (FCMs) provide a rapid and efficient approach for system modeling and simulation. The literature demonstrates numerous successful applications of FCMs in identifying failure modes. The standard process of failure mode identification using FCMs involves monitoring crucial concept/node values for excesses. Threshold functions are used to limit the value of nodes within a pre-specified range, which is usually [0, 1] or [-1, + 1]. However, traditional FCMs using the tanh threshold function possess two crucial drawbacks for this particular.Purpose(i) a tendency to reduce the values of state vector components, and (ii) the potential inability to reach a limit state with clearly identifiable failure states. The reason for this is the inherent mathematical nature of the tanh function in being asymptotic to the horizontal line demarcating the edge of the specified range. To overcome these limitations, this paper introduces a novel modified tanh threshold function that effectively addresses both issues.

In the unsupervised domain adaptation (UDA) (Akada et al. Self-supervised learning of domain invariant features for depth estimation, in: 2022 IEEE/CVF winter conference on applications of computer vision (WACV), pp 3377–3387 (2022). 10.1109/WACV51458.2022.00107) depth estimation task, a new adaptive approach is to use the bidirectional transformation network to transfer the style between the target and source domain inputs, and then train the depth estimation network in their respective domains. However, the domain adaptation process and the style transfer may result in defects and biases, often leading to depth holes and instance edge depth missing in the target domain’s depth output. To address these issues, We propose a training network that has been improved in terms of model structure and supervision constraints. First, we introduce a edge-guided self-attention mechanism in the task network of each domain to enhance the network’s attention to high-frequency edge features, maintain clear boundaries and fill in missing areas of depth. Furthermore, we utilize an edge detection algorithm to extract edge features from the input of the target domain. Then we establish edge consistency constraints between inter-domain entities in order to narrow the gap between domains and make domain-to-domain transfers easier. Our experimental demonstrate that our proposed method effectively solve the aforementioned problem, resulting in a higher quality depth map and outperforming existing state-of-the-art methods.

Image anomaly detection and localization perform not only image-level anomaly classification but also locate pixel-level anomaly regions. Recently, it has received much research attention due to its wide application in various fields. This paper proposes ProtoAD, a prototype-based neural network for image anomaly detection and localization. First, the patch features of normal images are extracted by a deep network pre-trained on nature images. Then, the prototypes of the normal patch features are learned by non-parametric clustering. Finally, we construct an image anomaly localization network (ProtoAD) by appending the feature extraction network with L2 feature normalization, a (1times 1) convolutional layer, a channel max-pooling, and a subtraction operation. We use the prototypes as the kernels of the (1times 1) convolutional layer; therefore, our neural network does not need a training phase and can conduct anomaly detection and localization in an end-to-end manner. Extensive experiments on two challenging industrial anomaly detection datasets, MVTec AD and BTAD, demonstrate that ProtoAD achieves competitive performance compared to the state-of-the-art methods with a higher inference speed. The code and pre-trained models are publicly available at https://github.com/98chao/ProtoAD.

Voice conversion (VC) is a task for changing the speech of a source speaker to the target voice while preserving linguistic information of the source speech. The existing VC methods typically use mel-spectrogram as both input and output, so a separate vocoder is required to transform mel-spectrogram into waveform. Therefore, the VC performance varies depending on the vocoder performance, and noisy speech can be generated due to problems such as train-test mismatch. In this paper, we propose a speech and fundamental frequency consistent raw audio voice conversion method called WaveVC. Unlike other methods, WaveVC does not require a separate vocoder and can perform VC directly on raw audio waveform using 1D convolution. This eliminates the issue of performance degradation caused by the train-test mismatch of the vocoder. In the training phase, WaveVC employs speech loss and F0 loss to preserve the content of the source speech and generate F0 consistent speech using the pre-trained networks. WaveVC is capable of converting voices while maintaining consistency in speech and fundamental frequency. In the test phase, the F0 feature of the source speech is concatenated with a content embedding vector to ensure the converted speech follows the fundamental frequency flow of the source speech. WaveVC achieves higher performances than baseline methods in both many-to-many VC and any-to-any VC. The converted samples are available online.

To address the challenges of the poor representation capability and low data utilization rate of end-to-end speech recognition models in deep learning, this study proposes an end-to-end speech recognition model based on multi-scale feature fusion and multi-view self-supervised learning (MM-ASR). It adopts a multi-task learning paradigm for training. The proposed method emphasizes the importance of inter-layer information within shared encoders, aiming to enhance the model’s characterization capability via the multi-scale feature fusion module. Moreover, we apply multi-view self-supervised learning to effectively exploit data information. Our approach is rigorously evaluated on the Aishell-1 dataset and further validated its effectiveness on the English corpus WSJ. The experimental results demonstrate a noteworthy 4.6(%) reduction in character error rate, indicating significantly improved speech recognition performance . These findings showcase the effectiveness and potential of our proposed MM-ASR model for end-to-end speech recognition tasks.

Few-shot class-incremental learning (FSCIL) struggles to incrementally recognize novel classes from few examples without catastrophic forgetting of old classes or overfitting to new classes. We propose TLCE, which ensembles multiple pre-trained models to improve separation of novel and old classes. Specifically, we use episodic training to map images from old classes to quasi-orthogonal prototypes, which minimizes interference between old and new classes. Then, we incorporate the use of ensembling diverse pre-trained models to further tackle the challenge of data imbalance and enhance adaptation to novel classes. Extensive experiments on various datasets demonstrate that our transfer learning ensemble approach outperforms state-of-the-art FSCIL methods.

Extracting relational triples from a piece of text is an essential task in knowledge graph construction. However, most existing methods either identify entities before predicting their relations, or detect relations before recognizing associated entities. This order may lead to error accumulation because once there is an error in the initial step, it will accumulate to subsequent steps. To solve this problem, we propose a parallel model for jointly extracting entities and relations, called PRE-Span, which consists of two mutually independent submodules. Specifically, candidate entities and relations are first generated by enumerating token sequences in sentences. Then, two independent submodules (Entity Extraction Module and Relation Detection Module) are designed to predict entities and relations. Finally, the predicted results of the two submodules are analyzed to select entities and relations, which are jointly decoded to obtain relational triples. The advantage of this method is that all triples can be extracted in just one step. Extensive experiments on the WebNLG*, NYT*, NYT and WebNLG datasets show that our model outperforms other baselines at 94.4%, 88.3%, 86.5% and 83.0%, respectively.

This study recommends a new time series forecasting model, namely ICEEMDAN - SVD - LSTM model, which coalesces Improved Complete Ensemble EMD with Adaptive Noise, Singular Value Decomposition and Long Short Term Memory network. It can be applied to analyse Non-linear and non-stationary data. The framework of this model is comprised of three levels, namely ICEEMDAN level, SVD level and LSTM level. The first level utilized ICEEMDAN to break up the series into some IMF components along with a residue. The SVD in the second level accounts for de-noising of every IMF component and residue. LSTM forecasts all the resultant IMF components and residue in third level. To obtain the forecasted values of the original data, the predictions of all IMF components and residue are added. The proposed model is contrasted with other extant ones, namely LSTM model, EMD - LSTM model, EEMD - LSTM model, CEEMDAN - LSTM model, EEMD - SVD - LSTM model, ICEEMDAN - LSTM model and CEEMDAN - SVD - LSTM model. The comparison bears witness to the potential of the recommended model over the traditional models.