Artificial Neural Networks (NNs) are unable to learn tasks continually using a single model, which leads to forgetting old knowledge, known as catastrophic forgetting. This is one of the shortcomings that usually plague intelligent systems based on NN models. Federated Learning (FL) is a decentralized approach to training machine learning models on multiple local clients without exchanging raw data. A paradigm that handles model learning in both settings, federated and continual, is known as Federated Continual Learning (FCL). In this work, we propose a novel FCL algorithm, called FedCluLearn, which uses a stream micro-cluster indexing scheme to deal with catastrophic forgetting. FedCluLearn interprets the federated training process as a stream clustering scenario. It stores statistics, similar to micro-clusters in stream clustering algorithms, about the learned concepts at the server and updates them at each training round to reflect the current local updates of the clients. FedCluLearn uses only active concepts in each training round to build the global model, meaning it temporarily forgets the knowledge that is not relevant to the current situation. In addition, the proposed algorithm is flexible in that it can consider the age of local updates to reflect the greater importance of more recent data. The proposed FCL approach has been benchmarked against three baseline algorithms by evaluating its performance in several control and real-world data experiments. The implementation of FedCluLearn and the experimental results are available at https://github.com/milenaangelova1/FedCluLearn.

Many machine learning models deployed on smart or edge devices experience a phase where there is a drop in their performance due to the arrival of data from new domains. This paper proposes a novel unsupervised domain adaptation algorithm called DIBCA++ to deal with such situations. The algorithm uses only the clusters’ mean, standard deviation, and size, which makes the proposed algorithm modest in terms of the required storage and computation. The study also presents the explainability aspect of the algorithm. DIBCA++ is compared with its predecessor, DIBCA, and its applicability and performance are studied and evaluated in two real-world scenarios. One is coping with the Global Navigation Satellite System activation problem from the smart logistics domain, while the other identifies different activities a person performs and deals with a human activity recognition task. Both scenarios involve time series data phenomena, i.e., DIBCA++ also contributes towards addressing the current gap regarding domain adaptation solutions for time series data. Based on the experimental results, DIBCA++ has improved performance compared to DIBCA. The DIBCA++ has performed better in all human activity recognition task experiments and 82.5% of experimental scenarios on the smart logistics use case. The results also showcase the need and benefit of personalizing the models using DIBCA++, along with the ability to transfer new knowledge between domains, leading to improved performance. The adapted source and target models have performed better in 70% and 80% of cases in an experimental scenario conducted on smart logistics. 

The AI marketplace ecosystem accelerates multiple modules of the AI engineering pipeline by fostering collaboration between stakeholders. However, marketplace collaborators often face a dilemma in striking a balance between sharing artifacts and protecting intellectual property (IP) rights. Thus, there is a need for a license management system within the AI marketplace to facilitate the exchange of artifacts in a trusted and secure manner. 

This work shares experiences while building such a license management system within the Bonseyes marketplace (BMP), a functional crowdsourcing AI marketplace that specializes in deploying real-time applications on edge devices. The BMP was developed, and its applicability is proven through the European H2020 project by a series of open calls and workshops, for gathering stakeholders and orchestrating the marketplace operations. 

The main contributions of this work are (i) implementation of an end-to-end license management system that deals with selecting license templates, license agreement interaction between seller and buyer, and the generation and enforcement of human- and machine-readable license files, and (ii) introduction of "Synchronization licenses'' concept from the music industry to the AI marketplace context where consumers acquire a license to integrate the artifact into another application, and a respective BMP use-case for collaborative AI engineering. 

Multi-user extended reality (XR) is an advanced technology that immerses users simultaneously into the meta-verse, aiming to provide a real-time interaction experience in different scenarios. However, different users' network conditions cause data transmission to be asynchronous. It forces users' behavior in the metaverse to go out of sync, seriously impacting the user experience. This paper presents a framework for human-centred synchronization of networks and metaverse applications, often denoted as Intelligent Realities, using Digital Twin (DT) technologies for the networks (network DTs) as well as for the individual users' XR parts (XR DTs). Our framework uses users' attention as a key input to dynamically allocate network resources and align content, thereby keeping widely distributed users in sync. To motivate this framework, we surveyed recent DT-based XR and network systems and found a significant gap: few studies work on synchronizing systems between network DTs and XR DTs. This highlights the need to explore integrated solutions for networked real-time multi-user interaction in the metaverse. 

Split Learning (SL) is an emerging decentralized paradigm that enables numerous participants, to train a deep neural network without disclosing sensitive information, such as patient data, in fields such as healthcare. In healthcare, SL enables distributed training across a variety of medical devices, hospitals, and organizations, improving model robustness while maintaining patient confidentiality. However, training models within SL is affected by data heterogeneity and sensitivity, and often requires more computational resources than an individual data provider can afford. This can result in significant model divergence and decreased performance due to differences in data distributions between various clients. To address this issue, we propose a framework that integrates fairness and adaptivity considerations, called ADF-SL. In particular, ADF-SL dynamically adjusts the total number of clients involved in model training and the number of iteration required to achieve convergence without compromising participant privacy. To evaluate performance, we compare the effectiveness of ADF-SL with that of the naive (Vanilla) SL approach, SplitFed and FairFed. Extensive experiments performed on time series electrocardiogram (ECG) databases (MITDB, SVDB, and INCARTDB) indicate that ADF-SL significantly outperforms the three existing algorithms that served as baselines. Compared to these baseline methods, ADF-SL accelerates model training on clients by up to 22.7%, 10.4%, and 5.8% compared to Vanilla SL, SplitFed, and FairFed, respectively, while maintaining model convergence and accuracy. Furthermore, the conducted ablation study has confirmed the importance of ADF-SL decay enrichment, which has outperformed non-decay ADF-SL for each used dataset by up to 15.8%, 43.9%, and 7.6%, respectively. 

In this study, we propose a novel neuro-symbolic approach to deal with the inherent ambiguity in image scene classification, combining the usage of pre-trained deep learning (DL) models with concepts from modal logic and evidence theory. The DL models are used to detect objects and estimate their depth in a set of labeled images. The obtained outputs are employed to form a dataset of instances characterizing the possible classes. Subsequently, a multi-valued mapping is defined between the data instances and the considered images resulting into each image being represented by the set of instances associated with it. The obtained mapping is utilized to infer necessity and possibility conditions of each class, or equivalently its upper (plausibility) and lower (belief) probabilities. Based on these interval evaluations, a rule-based and a score-based classifiers are built. The overall method is explainable and directly interpretable, robust to data scarcity and data imbalance. The presented framework is studied and evaluated on an abandoned bag detection use case. 

Extended reality (XR) technologies are increasingly being used in different application areas. One such area is for healthcare, which has seen significant developments over the last few years. However, its use for healthcare education is still in its infancy. This paper presents a case study, which explores the use of virtual reality (VR) technology in the healthcare domain. In particular, an application targeting education of various conditions healthcare providers might meet in older adult care is evaluated using different subjective evaluations methodologies, with nursing students and professional healthcare staff. The overall results show promising directions and use of new technology applications in this domain, but also highlights some of the potential problems with its adoption. 

Biosignal representation learning (BRL) plays a crucial role in emotion recognition for game users (ERGU). Unsupervised BRL has garnered attention considering the difficulty in obtaining ground truth emotion labels from game users. However, unsupervised BRL in ERGU faces challenges, including overfitting caused by limited data and performance degradation due to unbalanced sample distributions. Faced with the above challenges, we propose a novel method of biosignal contrastive representation learning (BCRL) for ERGU, which not only serves as a unified representation learning approach applicable to various modalities of biosignals but also derives generalized biosignals representations suitable for different downstream tasks. Specifically, we solve the overfitting by introducing perturbations at the embedding layer based on the projected gradient descent (PGD) adversarial attacks and develop the sample balancing strategy (SBS) to mitigate the negative impact of the unbalanced sample on the performance. Further, we have conducted comprehensive validation experiments on the public dataset, yielding the following key observations: first BCRL outperforms all other methods, achieving average accuracies of 76.67%, 71.83%, and 63.58% in 1D-2 C Valence, 1D-2 C Arousal, and 2D-4 C Valence/Arousal, respectively; second, the ablation study shows that both the PGD module (+7.58% in accuracy on average) and the SBS module (+14.60% in accuracy on average) have a positive effect on the performance of different classifications; third, BCRL model exhibits the certain generalization ability across the different games, subjects and classifiers.

Biosignal sequence real-time prediction (BSRP) is essential for predicting the future emotional experience of game users. However, BSRP for game users faces challenges, including poor real-time performance and limited feature fusion dimensions.

To address these issues, we proposed a method for BSRP based on the features fusion of Local-Global and Time-Frequency domain (LGTF) for game users, which integrates real-time capabilities with multi-dimensional features fusion. Specifically, LGTF meets real-time requirements and achieves the features fusion of Local-Global through multi-channel synchronized adaptive convolution. In addition, LGTF implements the features fusion of inter- and intra-band in the frequency domain and the features fusion of time-frequency domain by incorporating the Self-Attention mechanism and Fourier Transform. Furthermore, we conducted comprehensive validation experiments on LGTF using the public dataset.

The results indicate that: 1) In the comparison study, LGTF outperformed other methods, achieving the lowest average MSE and MAE values across different prediction lengths of 0.61 and 0.47, respectively. 2) Ablation studies revealed that the addition of time-frequency domain feature fusion (TF) and local-global feature fusion (LG) both have the positive effect on the prediction performance, reducing the average MSE by 0.11 and 0.09, respectively. 3) Generalization study shows that LGTF exhibits stable performance and generalization across different subjects and shows performance advantages in specific game scenarios. 4) Time performance analysis suggests LGTF has the real-time performance.5) Case study demonstrates that LGTF is practical for predicting game users' future emotions and enhancing their emotional experiences.

Federated learning (FL), a decentralized machine learning framework that allows edge devices (i.e., clients) to train a global model while preserving data/client privacy, has become increasingly popular recently. In FL, a shared global model is built by aggregating the updated parameters in a distributed manner. To incentivize data owners to participate in FL, it is essential for service providers to fairly evaluate the contribution of each data owner to the shared model during the learning process. To the best of our knowledge, most existing solutions are resource-demanding and usually run as an additional evaluation procedure. The latter produces an expensive computational cost for large data owners. In this paper, we present simple and effective FL solutions that show how the clients’ behavior can be evaluated during the training process with respect to reliability, and this is demonstrated for two existing FL models, Cluster Analysis-based Federated Learning (CA-FL) and Group-Personalized FL (GP-FL), respectively. In the former model, CA-FL, the frequency of each client to be selected as a cluster representative and in that way to be involved in the building of the shared model is assessed. This can eventually be considered as a measure of the respective client data reliability. In the latter model, GP-FL, we calculate how many times each client changes a cluster it belongs to during FL training, which can be interpreted as a measure of the client's unstable behavior, i.e., it can be considered as not very reliable. We validate our FL approaches on three LEAF datasets and benchmark their performance to two baseline contribution evaluation approaches. The experimental results demonstrate that by applying the two FL models we are able to get robust evaluations of clients’ behavior during the training process. These evaluations can be used for further studying, comparing, understanding, and eventually predicting clients’ contributions to the shared global model.