The paper describes a new type of evolving connectionist systems (ECOS) called evolving spatio-temporal data machines based on neuromorphic, brain-like information processing principles (eSTDM). These are multi-modular computer systems designed to deal with large and fast spatio/spectro temporal data using spiking neural networks (SNN) as major processing modules. ECOS and eSTDM in particular can learn incrementally from data streams, can include 'on the fly' new input variables, new output class labels or regression outputs, can continuously adapt their structure and functionality, can be visualised and interpreted for new knowledge discovery and for a better understanding of the data and the processes that generated it. eSTDM can be used for early event prediction due to the ability of the SNN to spike early, before whole input vectors (they were trained on) are presented. A framework for building eSTDM called NeuCube along with a design methodology for building eSTDM using this is presented. The implementation of this framework in MATLAB, Java, and PyNN (Python) is presented. The latter facilitates the use of neuromorphic hardware platforms to run the eSTDM. Selected examples are given of eSTDM for pattern recognition and early event prediction on EEG data, fMRI data, multisensory seismic data, ecological data, climate data, audio-visual data. Future directions are discussed, including extension of the NeuCube framework for building neurogenetic eSTDM and also new applications of eSTDM.
Document classification aims to assign one or more classes to a document for ease of management by understanding the content of a document. Hierarchical attention network (HAN) has been showed effective to classify documents that are ambiguous. HAN parses information-intense documents into slices (i.e., words and sentences) such that each slice can be learned separately and in parallel before assigning the classes. However, introducing hierarchical attention approach leads to the redundancy of training parameters which is prone to overfitting. To mitigate the concern of overfitting, we propose a variant of hierarchical attention network using adversarial and virtual adversarial perturbations in 1) word representation, 2) sentence representation and 3) both word and sentence representations. The proposed variant is tested on eight publicly available datasets. The results show that the proposed variant outperforms the hierarchical attention network with and without using random perturbation. More importantly, the proposed variant achieves state-of-the-art performance on multiple benchmark datasets. Visualizations and analysis are provided to show that perturbation can effectively alleviate the overfitting issue and improve the performance of hierarchical attention network.
Paraphrase identification serves as an important topic in natural language processing while sequence alignment and matching underlie the principle of this task. Traditional alignment methods take advantage of attention mechanism. Attention mechanism, i.e. weighting technique, could pick out the most similar/dissimilar parts, but is weak in modeling the aligned unmatched parts, which are the crucial evidence to identify paraphrases. In this paper, we empower neural architecture with Hungarian algorithm to extract the aligned unmatched parts. Specifically, first, our model applies BiLSTM/BERT to encode the input sentences into hidden representations. Then, Hungarian layer leverages the hidden representations to extract the aligned unmatched parts. Last, we apply cosine similarity to metric the aligned unmatched parts for a final discrimination. Extensive experiments show that our model outperforms other baselines, substantially and significantly.
In this paper, we extend our previous work on the Enhanced Fuzzy Min-Max (EFMM) neural network by introducing a new hyperbox selection rule and a pruning strategy to reduce network complexity and improve classification performance. Specifically, a new k-nearest hyperbox expansion rule (for selection of a new winning hyperbox) is first introduced to reduce the network complexity by avoiding the creation of too many small hyperboxes within the vicinity of the winning hyperbox. A pruning strategy is then deployed to further reduce the network complexity in the presence of noisy data. The effectiveness of the proposed network is evaluated using a number of benchmark data sets. The results compare favorably with those from other related models. The findings indicate that the newly introduced hyperbox winner selection rule coupled with the pruning strategy are useful for undertaking pattern classification problems.
This study aims to explore an effective method to evaluate spatial cognitive ability, which can effectively extract and classify the feature of EEG signals collected from subjects participating in the virtual reality (VR) environment; and evaluate the training effect objectively and quantitatively to ensure the objectivity and accuracy of spatial cognition evaluation, according to the classification results. Therefore, a multi-dimensional conditional mutual information (MCMI) method is proposed, which could calculate the coupling strength of two channels considering the influence of other channels. The coupled characteristics of the multi-frequency combination were transformed into multi-spectral images, and the image data were classified employing the convolutional neural networks (CNN) model. The experimental results showed that the multi-spectral image transform features based on MCMI are better in classification than other methods, and among the classification results of six band combinations, the best classification accuracy of Beta1-Beta2-Gamma combination is 98.3%. The MCMI characteristics on the Beta1-Beta2-Gamma band combination can be a biological marker for the evaluation of spatial cognition. The proposed feature extraction method based on MCMI provides a new perspective for spatial cognitive ability assessment and analysis.
Adaptive Resonance Theory (ART) is one of the successful approaches to resolving "the plasticity-stability dilemma" in neural networks, and its supervised learning model called ARTMAP is a powerful tool for classification. Among several improvements, such as Fuzzy or Gaussian based models, the state of art model is Bayesian based one, while solving the drawbacks of others. However, it is known that the Bayesian approach for the high dimensional and a large number of data requires high computational cost, and the covariance matrix in likelihood becomes unstable. This paper introduces Kernel Bayesian ART (KBA) and ARTMAP (KBAM) by integrating Kernel Bayes' Rule (KBR) and Correntropy Induced Metric (CIM) to Bayesian ART (BA) and ARTMAP (BAM), respectively, while maintaining the properties of BA and BAM. The kernel frameworks in KBA and KBAM are able to avoid the curse of dimensionality. In addition, the covariance-free Bayesian computation by KBR provides the efficient and stable computational capability to KBA and KBAM. Furthermore, Correntropy-based similarity measurement allows improving the noise reduction ability even in the high dimensional space. The simulation experiments show that KBA performs an outstanding self-organizing capability than BA, and KBAM provides the superior classification ability than BAM, respectively.
Deep learning-based models have achieved significant success in detecting cardiac arrhythmia by analyzing ECG signals to categorize patient heartbeats. To improve the performance of such models, we have developed a novel hybrid hierarchical attention-based bidirectional recurrent neural network with dilated CNN (HARDC) method for arrhythmia classification. This solves problems that arise when traditional dilated convolutional neural network (CNN) models disregard the correlation between contexts and gradient dispersion. The proposed HARDC fully exploits the dilated CNN and bidirectional recurrent neural network unit (BiGRU-BiLSTM) architecture to generate fusion features. As a result of incorporating both local and global feature information and an attention mechanism, the model's performance for prediction is improved. By combining the fusion features with a dilated CNN and a hierarchical attention mechanism, the trained HARDC model showed significantly improved classification results and interpretability of feature extraction on the PhysioNet 2017 challenge dataset. Sequential Z-Score normalization, filtering, denoising, and segmentation are used to prepare the raw data for analysis. CGAN (Conditional Generative Adversarial Network) is then used to generate synthetic signals from the processed data. The experimental results demonstrate that the proposed HARDC model significantly outperforms other existing models, achieving an accuracy of 99.60%, F1 score of 98.21%, a precision of 97.66%, and recall of 99.60% using MIT-BIH generated ECG. In addition, this approach significantly reduces run time when using dilated CNN compared to normal convolution. Overall, this hybrid model demonstrates an innovative and cost-effective strategy for ECG signal compression and high-performance ECG recognition. Our results indicate that an automated and highly computed method to classify multiple types of arrhythmia signals holds considerable promise.
Few-shot learning aims to train a model with a limited number of base class samples to classify the novel class samples. However, to attain generalization with a limited number of samples is not a trivial task. This paper proposed a novel few-shot learning approach named Self-supervised Contrastive Learning (SCL) that enriched the model representation with multiple self-supervision objectives. Given the base class samples, the model is trained with the base class loss. Subsequently, contrastive-based self-supervision is introduced to minimize the distance between each training sample with their augmented variants to improve the sample discrimination. To recognize the distant sample, rotation-based self-supervision is proposed to enable the model to learn to recognize the rotation degree of the samples for better sample diversity. The multitask environment is introduced where each training sample is assigned with two class labels: base class label and rotation class label. Complex augmentation is put forth to help the model learn a deeper understanding of the object. The image structure of the training samples are augmented independent of the base class information. The proposed SCL is trained to minimize the base class loss, contrastive distance loss, and rotation class loss simultaneously to learn the generic features and improve the novel class performance. With the multiple self-supervision objectives, the proposed SCL outperforms state-of-the-art few-shot approaches on few-shot image classification benchmark datasets.