ProfessorNikola Kasabov

Quantum algorithms have become a popular research domain in recent times for discovering quantum-enhanced solutions in machine learning applications. Quantum kernels are one of the directions that establish such quantum-enhanced solutions to some extent. This work presents a detailed analysis of the quantum kernel approach leveraging feature maps and relevant hyperparameters to develop enhanced quantum kernels. The study includes a new high-order feature map and assesses five existing state-of-the-art feature maps for enhanced quantum kernel classifiers. Additionally, the significance of the rotational factor as a hyperparameter is highlighted for improving kernel performance. Also, it is analyzed whether different hyperparameter-tuned feature maps can lead to enhanced decision boundaries, demonstrating kernel expressivity. The analysis is undertaken on classification tasks using four different nonlinear datasets of distinct complexity. Comparative evaluations are also made with traditional machine learning models—Support Vector Machines (Linear and RBF), Naïve Bayes, Linear Discriminant Analysis, Decision Tree, Random Forest, Adaptive Boosting, and MLP. Overall, the study demonstrates that a well-tuned quantum feature map can enhance the generalization ability of quantum kernels, making them more effective for broader quantum-enhanced machine learning applications.

Agentic AI is shifting AI applications from passive model inference to goal-driven, tool-using, and collaborative autonomous systems. Yet, current deployments remain concentrated in data centers or powerful personal devices. This paper provides a research roadmap for enabling large-scale, enterprise-oriented agentic AI across the computing continuum, from cloud to edge, IoT, and emerging hardware platforms, where heterogeneity, mobility, energy constraints, and governance requirements fundamentally reshape agent design and operation. We argue for a continuum-native agent paradigm that decomposes agents into intelligence, persona, and memory, allowing their independent placement, migration, and replication. Building on this abstraction, we structure the open research space into three pillars (agent operation, agent connectivity, and agent trustworthiness) and elaborate on potential research directions and challenges. Finally, we outline evaluation directions grounded in representative industrial use cases and benchmarks. The roadmap aims to guide research and development toward enabling the deployment of agentic AI as first-class workloads on the cloud-edge continuum.

This paper introduces a biomimetic framework and novel brain-inspired AI (BIAI) models based on spiking neural networks (SNNs) for emotional state recognition from audio (speech), visual (face), and integrated multimodal audio-visual data. The developed framework, named eXCube2, uses a three-dimensional SNN architecture NeuCube that is spatially structured according to a human brain template. The BIAI models developed in eXCube2 are trainable on spatio- and spectro-temporal data using brain-inspired learning rules. Such models are explainable in terms of revealing patterns in data and are adaptable to new data. The eXCube2 models are implemented as software systems and tested on speech and video data of subjects expressing emotional states. The use of a brain template for the SNN structure enables brain-inspired tonotopic and stereo mapping of audio inputs, topographic mapping of visual data, and the combined use of both modalities. This novel approach brings AI-based emotional state recognition closer to human perception, provides a better explainability and adaptability than existing AI systems. It also results in a higher or competitive accuracy, even though this was not the main goal here. This is demonstrated through experiments on benchmark datasets, achieving classification accuracy above 80% on single-modality data and 88.9% when multimodal audio-visual data are used, and a "don't know" output is introduced. The paper further discusses possible applications of the proposed eXCube2 framework to other audio, visual, and audio-visual data for solving challenging problems, such as recognizing emotional states of people from different origins; brain state diagnosis (e.g., Parkinson's disease, Alzheimer's disease, ADHD, dementia); measuring response to treatment over time; evaluating satisfaction responses from online clients; cognitive robotics; human-robot interaction; chatbots; and interactive computer games. The SNN-based implementation of BIAI also enables the use of neuromorphic chips and platforms, leading to reduced power consumption, smaller device size, higher performance accuracy, and improved adaptability and explainability. This research shows a step toward building brain-inspired AI systems.

This paper presents a novel spiking neural network (SNN) framework that integrates unsupervised and supervised learning to improve classification accuracy while reducing model complexity. Spatio-temporal data are modelled using a three-dimensional SNN reservoir trained via spike-timing-dependent plasticity (STDP). A dynamic clustering method is applied during STDP to analyse evolving spike patterns and prune less informative connections, resulting in a sparse, functionally rich reservoir. We propose a new supervised learning model, Neuron-Stability Weighted Dynamic Evolving Spiking Neural Network (NSW-deSNN), which extends the traditional deSNN by incorporating a stability-based weighting term. This term, derived from the entropy of neuron-cluster memberships, prioritizes contributions from stable, informative neurons when forming output connections. Tested on fMRI data related to a cognitive task (reading sentences with different polarities), the proposed approach demonstrates improved accuracy and efficiency over conventional fully connected SNN models, highlighting its potential for learning meaningful spatio-temporal patterns from brain data.

Image fusion aims to produce high-quality fused images by merging complementary information from source images, further boosting the performance of subsequent tasks. Recently, the text modality has been integrated into image fusion frameworks to improve the semantic richness of fused images. However, existing approaches struggle with semantic alignment and fail to capture complex cross-modal relationships, ultimately leading to suboptimal fusion quality in complex scenes. To tackle the aforementioned problems, we introduce an innovative fusion framework that leverages text-guided semantic perception to enhance infrared and visible image fusion, termed as TSPFusion. Unlike previous works that utilize language models to generate scene semantic description, TSPFusion leverages segmentation annotations from downstream tasks to generate accurate object-level text prompts. The proposed framework first adopts a dual-branch backbone to independently encode modality-specific information from infrared and visible inputs. Then, we introduce a Semantic Correspondence Alignment Module (SCAM) to align these features with text prompts, which integrates a pre-trained CLIP model and a prompt optimization mechanism to refine textual representations. Finally, a Prompt-guided Semantic Modulation Module performs hierarchical modulation during decoding. At each decoding layer, the original and refined text embeddings are employed by PSMM to modulate encoder and decoder features to ensure structural integrity and semantic consistency. Extensive experiments have validated the superiority of TSPFusion over state-of-the-art fusion approaches, delivering consistently better performance on various benchmark tasks.

The study introduces a hybrid computational framework that combines neuro-inspired information processing using spiking neural networks (SNNs) and quantum information processing using quantum kernels to develop quantum-enhanced machine learning models for spatio-temporal data, demonstrated through the classification of EEG data as a case study. In the proposed SNN-quantum computation (SNN-QC) framework, SNN with spike time information representation is employed to learn spatio-temporal interactions (EEG recorded from multiple channels over time). Frequency-based (rate-based) information as spike frequency state vectors are extracted from the SNN and classified using a quantum classifier. In the latter part, we use the quantum kernel approach utilising feature maps for classification tasks. The proposed SNN-QC is demonstrated on a benchmark EEG dataset to classify three distinct wrist movement tasks in six binary classification setups as a proof of concept. We introduce a novel high-order nonlinear feature map that demonstrates improved performance over state-of-the-art feature maps and several machine learning methods across most of the tasks studied. Furthermore, the role of hyperparameters for enhanced feature maps is also highlighted. The performance of SNN-QC is evaluated using statistical metrics and cross-validation techniques, demonstrating its efficacy across multiple binary classifiers. Quantum hardware validation is conducted using both a superconducting IBM-QPU and a high-fidelity noisy simulation that replicates a real QPU. Furthermore, the results demonstrate that the SNN-QC outperforms models that use statistical features rather than features extracted from the SNN, as the SNN accounts for the temporal interaction between the spatio-temporal input variables. Finally, we conclude that the SNN-QC offers a potential pathway for developing more accurate neuromorphic-quantum enhanced systems that are both energy-efficient and biologically-inspired, well-suited for dealing with spatio-temporal data.

The process of fusing two complementary data, panchromatic and multispectral images, to create high-resolution multispectral (HRMS) images is known as pansharpening. Combining detail injection (DI) methods with convolutional neural networks (CNN) for improved HRMS image fusion quality is a research hotspot due to their interpretability and large-scale data processing capabilities, respectively. Nevertheless, the current hybrid models typically concatenate CNN and traditional techniques, limiting the ability to utilize the benefits of both approaches. This paper presents a new hybrid network, multispectral pansharpening method based on CNN-DI network with mixture of experts (CDN-MoE), using detail injection theory to design a deep learning framework. Specifically, we first create the mixture of detail inject experts network (MoDIE-Net) that mixes training pairs of full- and reduced-resolution images to enhance model generalization. Next, the adaptive correlation residual network (ACR-Net) is suggested to find the correlation between the spectral and spatial features of the source images. Finally, the global information injection network (GII-Net) is established to strengthen the accuracy of fusion results by integrating the context of input images. Additionally, to reduce the loss of spectral features during the upsampling process, the spectral reconstruction network (SR-Net) is proposed. We perform both qualitative and quantitative experiments on the GaoFen-2, IKONOS, and WorldView-2 datasets at various resolutions. Our approach has advantages over other SOTA pansharpening methods currently available in terms of visual effects and objective metrics.

The study introduces a novel computational framework combining neuro-inspired information through spiking neural networks (SNNs) and quantum information using quantum kernels to develop quantum machine learning models. The framework uses a 3D brain-inspired SNN model to learn spatiotemporal EEG signals through their dynamic interaction. Spike frequency state vectors are extracted from a 3D SNN and a quantum kernel classifier is used to classify them into predefined classes. The study also proposes a new embedding function to enhance the performance of the quantum kernel classifier. The performance of the model is evaluated using statistical metrics and cross-validation techniques, demonstrating its superior efficacy on multiple state-of-the-art embedding functions and classical baseline classifiers. The results indicate a clear advantage of having an integrated framework of an SNN and a quantum kernel, which overcomes the limitations of existing classifiers in dealing with spatiotemporal data and enhances the performance of existing SNNs by improving the classification of their complex internal states.

Havening is an innovative psychosensory intervention that uses touch to facilitate recovery from psychological trauma. Whilst Havening is commonly employed by practitioners worldwide, only a few empirical studies have investigated efficacy or mechanism. The current study applies explainable machine learning methods to brain data to better understand mechanisms underpinning the role of Havening touch in recovery from trauma. Participants (n = 27) who had experienced an event that caused persistent psychological distress underwent a single Havening session that either did (H+, n = 15) or did not (H−, n = 12) include a touch component. Resting-state electroencephalography (EEG) data was recorded before (T1) and after (T2) the intervention. Two H+ subgroups were formed based on self-reported-response to the intervention. A recently developed brain-inspired machine-learning model, the Spiking Neural Network (SNN) was applied to compare groups and time points on functional connectivity. Results suggest region-specific reductions (left temporal, left frontocentral) in connection weights following Havening touch. Differentiation of H+ and H− was more accurate at T2 than T1; and in H+, brain states at T2 were more accurately classified than at T1, particularly for participants who had a greater response to the intervention. Findings support the SNN in distinguishing brain states associated with response to Havening. Reduction in left temporal connectivity may reflect downregulation of anterior temporal lobe activity, downstream from the amygdala.

Current technological solutions for Brain-machine Interfaces (BMI) achieve reasonable accuracy, but most systems are large in size, power consuming and not auto-adaptive. This work addresses the question whether current neuromorphic technologies could resolve these problems? The paper proposes a novel neuromorphic framework of a BMI system for prosthetics control via decoding Electro Cortico-Graphic (ECoG) brain signals. It includes a three-dimensional spike timing neural network (3D-SNN) for brain signals features extraction and an on-line trainable recurrent reservoir structure (Echo state network (ESN)) for Motor Control Decoding (MCD). A software system, written in Python using NEST Simulator SNN library is described. It is able to adapt continuously in real time in supervised or unsupervised mode. The proposed approach was tested on several experimental data sets acquired from a tetraplegic person. First simulation results are encouraging, showing also the need for a further improvement via multiple hyper-parameters tuning. Its future implementation on a neuromorphic hardware platform that is smaller in size and significantly less power consuming is discussed too.

Associative memories (AM) are at the core of human intelligence and learning systems. While there have been some neural network AM developed for vector-based data such as images, current machine learning methods, including deep neural networks, do not allow for training a model on time series data and recalling it on a subset of variables measured over a shorter time window. They also do not support further incremental training of the model on new temporal data and new variables. This paper introduces a new framework and method for the creation of evolving predictive associative memories for time series, abbreviated here as ePAMeT. The method is based on spiking neural networks (SNN). ePAMeT introduces significant adaptability in handling time series data with reduced or newly introduced features. This model maintains high accuracy and explainability, offering substantial improvements over traditional methods in dynamic and uncertain environments. First, an SNN model is trained on multiple time series using all available variables measured at a full-time length, and then the model is recalled on subsets of variables at a shorter time measurement without compromising predictive accuracy. Using a shorter time for recall makes early prediction of events possible. The SNN model can be further adapted/evolved on new data without pre-training the model on the old data, even using new variables. This is possible due to the evolving connectivity of the SNN model. A dynamic graph is extracted from the SNN model to capture dynamic interactions between the used temporal variables at any time during the evolution of the model, which constitutes strong explainability and a generation of new knowledge. The method is illustrated on original financial time series data, but it is applicable to many other domain areas as discussed. The proposed method has advantages over traditional machine learning methods in terms of evolvability, explainability, knowledge discovery, and using partial information of both the number of variables and their time length for the recall of the model on new data. The proposed framework opens the field for creating new types of evolvable time series prediction models. Future developments are discussed.

The identification of reliable biomarkers predicting outcomes in ultra-high risk (UHR) psychosis remains a vital challenge in preventive psychiatry. While genetic factors are implicated in psychosis risk, specific markers predicting social and cognitive trajectories have remained elusive. In this 24-month longitudinal study of UHR and Healthy Control (HC) participants, we investigated for the first time the predictive relationship between key immune-related genes, inflammatory response genes, stress response genes, and neurological-related genes with social and cognitive outcomes. Participants underwent comprehensive social and cognitive assessments at baseline and three 6-month intervals. We employed a brain-inspired Spiking Neural Network (SNN) model to map gene-behavior interactions and to identify key genetic predictors that influence both social and cognitive functioning. Additionally, we explore subgroup differences within the UHR population to further understand psychosis risk.

BACKGROUND: Identifying biomarkers that predict social and cognitive outcomes in individuals at ultra-high risk (UHR) for psychosis remains a key challenge in preventive psychiatry. While genetic factors contribute to psychosis vulnerability, specific markers that predict individual trajectories of functional decline or resilience are still unclear. METHODS: In a 24-month longitudinal study involving UHR (n = 45) and healthy control participants (n = 54), we investigated for the first time the predictive causal relationship between key immunological genes (FABP5 family and immunoglobulins) and social-cognitive outcomes. Participants completed comprehensive assessments at baseline and four 6-month intervals. We used regression modelling and dynamic Bayesian network analysis to identify predictive relationships between gene expression and behavioral outcomes over time. RESULTS: FABP5 family genes (FABP5P1, FABP5P11, FABP5P9) significantly predicted verbal memory (β = 0.233, p = 0.002); working memory (β = 0.225, p = 0.004), and social skills (β =-0·190, p < 0.029), respectively, at 24 months in the UHR group. Immunoglobulin-related genes showed distinct effects: FCGR2B predicted object recognition ability (β = 0.233, p = 0.014), while GOT2 inversely predicted planning ability (β = -0.147, p = 0.067). Network analysis revealed UHR-specific temporal dependencies absent in controls, with FCGRT emerging as a central node linking genetic markers to changes in processing speed and perceptual closure. CONCLUSIONS: This study provides the first evidence that FABP5 and immunoglobulin-related genetic markers can predict social-cognitive trajectories in individuals at risk for psychosis. These findings support the use of genetic profiling for early identification and highlight new opportunities for personalized preventive strategies in psychiatry.

This study focuses on the integration of the Izhikevich neuron model into the NeuCube framework to improve classification accuracy in longitudinal data. NeuCube is a reservoir spiking neural network architecture and offers advanced features such as spatial neuron connectivity and biologically plausible learning, making it suitable for complex machine learning tasks. For extracting meaningful features from the spiking activity of NeuCube, we evaluated various sampling methods on the classification performance of control and UHR samples from the LYRIKS RNA sequencing dataset. Our findings demonstrate that feature extraction using temporal binning, particularly with 10 bins, yields the highest accuracy across neuron types by effectively capturing fine-grained temporal dynamics. Intrinsic bursting neurons showed superior accuracy across most sampling methods, underscoring their versatility in information transmission.