Cognitive Computation最新文献

Background

Federated Continual Learning (FCL) involves learning from distributed data on edge devices with incremental knowledge. However, current FCL methods struggle to retain long-term memories on the server.

Method

In this paper, we introduce a method called Fisher INformation Accumulation Learning (FINAL) to address catastrophic forgetting in FCL. First, we accumulate a global Fisher with a federated Fisher information matrix formed from clients task by task to remember long-term knowledge. Second, we present a novel multi-node collaborative integration strategy to assemble the federated Fisher, which reveals the task-specific co-importance of parameters among clients. Finally, we raise a Fisher balancing method to combine the global Fisher and federated Fisher, avoiding neglecting new learning or causing catastrophic forgetting.

Results

We conducted evaluations on four FCL datasets, and the findings demonstrate that the proposed FINAL effectively maintains long-term knowledge on the server.

Conclusions

The exceptional performance of this method indicates its significant value for future FCL research.

Smart healthcare systems promise significant benefits for fast and accurate medical decisions. However, working with personal health data presents new privacy issues and constraints that must be solved from a cybersecurity perspective. Edge intelligence-enabled federated learning is a new scheme that utilises decentralised computing that allows data analytics to be carried out at the edge of a network, enhancing data privacy. However, this scheme suffers from privacy attacks, including inference, free-riding, and man-in-the-middle attacks, especially with serverless computing for allocating resources to user needs. Edge intelligence-enabled federated learning requires client data insertion and deletion to authenticate genuine clients and a serverless computing capability to ensure the security of collaborative machine learning models. This work introduces a serverless privacy edge intelligence-based federated learning (SPEI-FL) framework to address these issues. SPEI-FL includes a federated edge aggregator and authentication method to improve the data privacy of federated learning and allow client adaptation and removal without impacting the overall learning processes. It also can classify intruders through serverless computing processes. The proposed framework was evaluated with the unstructured COVID-19 medical chest x-rays and MNIST digit datasets, and the structured BoT-IoT dataset. The performance of the framework is comparable with existing authentication methods and reported a higher accuracy than comparable methods (approximately 90% as compared with the 81% reported by peer methods). The proposed authentication method prevents the exposure of sensitive patient information during medical device authentication and would become the cornerstone of the next generation of medical security with serverless computing.

The burgeoning practice of unauthorized acquisition and utilization of personal textual data (e.g., social media comments and search histories) by certain entities has become a discernible trend. To uphold data protection regulations such as the Asia–Pacific Privacy Initiative (APPI) and to identify instances of unpermitted exploitation of personal data, we propose a novel and efficient audit framework that helps users conduct cognitive analysis to determine if their textual data was used for data augmentation. and training a discriminative model. In particular, we focus on auditing models that use BERT as the backbone for discriminating text and are at the core of popular online services. We first propose an accumulated discrepancy score, which involves not only the response of the target model to the auditing sample but also the responses between pre-trained and finetuned models, to identify membership. We implement two types of audit methods (i.e., sample-level and user-level) according to our framework and conduct comprehensive experiments on two downstream applications to evaluate the performance. The experimental results demonstrate that our sample-level auditing achieves an AUC of 89.7% and an accuracy of 83%, whereas the user-level method can audit membership with an AUC of 89.7% and an accuracy of 88%. Additionally, we undertake an analysis of how augmentation methods impact auditing performance and expound upon the underlying reasons for these observations.

In medical imaging, traditional methods have long been relied upon. However, the integration of Generative Adversarial Networks (GANs) has sparked a paradigm shift, ushering in a new era of innovation. Our comprehensive investigation explores the groundbreaking impact of GANs on medical imaging, examining the evolution from traditional techniques to GAN-driven approaches. Through meticulous analysis, we dissect various aspects of GANs, encompassing their taxonomy, historical progression, and diverse iterations such as Self-Attention GANs (SAGAN), Conditional GANs, and Progressive Growing GANs (PGGAN). Complemented by a practical case study, we scrutinize the extensive applications of GANs, spanning image generation, reconstruction, enhancement, segmentation, and super-resolution. Despite promising prospects, enduring challenges including data scarcity, interpretability issues, and ethical concerns persist. Looking ahead, we anticipate advancements in personalized and pathological image generation, cross-modal synthesis, real-time interactive image generation, and enhanced anomaly detection. Through this review, we underscore the transformative potential of GANs in reshaping medical imaging practices, while also outlining avenues for future research endeavors.

Current brain science reveals that the connectivity patterns of the human brain are constantly changing when performing different tasks. Thus, brain effective connectivity networks based on non-stationary assumption can describe such neurodynamics better than the ones based on stationary assumption. However, existing methods for inferring non-stationary brain effective connectivity networks are committed to estimating the change points and network structures simultaneously. It is even worse that these methods will inevitably focus on one part of the estimation process and lead to the deviation of the results obtained by the other part. Then, the construction results of non-stationary brain effective connectivity networks cannot accurately reflect the real brain dynamics. In this paper, a novel approach to constructing non-stationary brain effective connectivity networks is proposed, namely CPD-NSL. It involves two stages including change point detection and network structure learning. In the first stage, the latent block model is used, and then the improved forward-backward search method is used to construct the stationary networks between adjacent change points in the network structure learning part. Finally, the constructed stationary networks are arranged in chronological order to obtain the final time-varying brain effective connectivity network. CPD-NSL is validated using simulated data as well as real fMRI data from HCP public datasets. The results show that CPD-NSL can restore the real network more accurately and consume less time. Experimental results on both simulated and real data demonstrate the effectiveness of the proposed method in constructing non-stationary state brain effective connectivity networks.

As is well known that time delays are inevitable in practice due to the finite switching speed of amplifiers and information transmission between neurons. So the study on the Hopf bifurcation of delayed neural networks has aroused extensive attention in recent years. However, it’s worth mentioning that only the communication delays between neurons were generally considered in most existing relevant literatures. Actually, it has been proven that a kind of so-called leakage delays cannot be ignored because the self-decay process of a neuron’s action potential is not instantaneous in hardware implementation of neural networks. Though leakage delays have been taken into account in a few more recent works concerning the Hopf bifurcation of fractional-order bidirectional associative memory neural networks, the addressed neural networks were low-dimension or the involved time delays were single. In this paper, we propose a five-neuron fractional-order bidirectional associative memory neural network model, which includes leakage delays and discrete communication delays to meet the characteristics of real neural networks better. Then we use the stability theory of fractional differential equations and Hopf bifurcation theory to investigate its dynamic behavior of Hopf bifurcation. The Hopf bifurcation of the proposed model are studied by taking the involved two different leakage delays as the bifurcation parameter respectively, and two kinds of sufficient conditions for Hopf bifurcation are obtained. A numerical example as well as its simulation plots and phase portraits are given at last. Our results indicate that a Hopf bifurcation rises near the zero equilibrium point when the leakage delay reaches its critical value which is given by an explicit formula. Particularly, the results of numerical simulations show that the leakage delay would narrow the stability region of the proposed system and make the Hopf bifurcation occur earlier.

The COVID-19 prediction process is more indispensable to handle the spread and death occurred rate because of COVID-19. However, early and precise prediction of COVID-19 is more difficult, because of different sizes and resolutions of input image. Thus, these challenges and problems experienced by traditional COVID-19 detection methods are considered as major motivation to develop SJHBO-based Deep Q Network. The classification issue of respiratory sound has perceived a great focus from the clinical scientists as well as the community of medical researcher in the previous year for the identification of COVID-19 disease. The major contribution of this research is to design an effectual COVID-19 detection model using devised SJHBO-based Deep Q Network. In this paper, the COVID-19 detection is carried out by the deep learning with optimization technique, namely Snake Jaya Honey Badger Optimization (SJHBO) algorithm-driven Deep Q Network. Here, the SJHBO algorithm is the incorporation of Jaya Honey Badger Optimization (JHBO) along with Snake optimization (SO). Here, the COVID-19 is detected by the Deep Q Network wherein the weights of Deep Q Network are tuned by the SJHBO algorithm. Moreover, JHBO is modelled by hybrids, which are the Jaya algorithm and Honey Badger Optimization (HBO) algorithm. Furthermore, the features, such as spectral contrast, Mel frequency cepstral coefficients (MFCC), empirical mode decomposition (EMD) algorithm, spectral flux, fast Fourier transform (FFT), spectral roll-off, spectral centroid, zero-crossing rate, root mean square energy, spectral bandwidth, spectral flatness, power spectral density, mobility complexity, fluctuation index and relative amplitude, are mined for enlightening the detection performance. The developed method realized the better performance based on the accuracy, sensitivity and specificity of 0.9511, 0.9506 and 0.9469. All test results are validated with the k-fold cross validation method in order to make an assessment of the generalizability of these results. Statistical analysis is performed to analyze the performance of the proposed method based on testing accuracy, sensitivity and specificity. Hence, this paper presents the newly devised SJHBO-based Deep Q-Net for COVID-19 detection. This research considers the audio samples as an input, which is acquired from the Coswara dataset. The SJHBO-based Deep Q network approach is developed for COVID-19 detection. The developed approach can be extended by including other hybrid optimization algorithms as well as other features that can be extracted for further improving the detection performance. The proposed COVID-19 detection method is useful in various applications, like medical and so on. Developed SJHBO-enabled Deep Q network for COVID-19 detection: An effective COVID-19 detection technique is introduced based on hybrid optimization–driven deep learning model. The Deep Q Network is used for detecting COVID-19, which classifies the feature vector