Digital accessibility revolves around the design and production of digital content and electronic devices easy to use and adapt to, meeting all audience regardless of physical or mental ability. While Digital accessibility is very important for ensuring that individuals with disabilities can fully participate in the digital space with adequate web adjustments, Artificial Intelligence (AI) has shown a huge potential in the advancement of digital accessibility. These two fields together have created an opportunity to solve the needs of individuals lacking equal access. AI has leveraged on non-visual modalities such as haptic, auditory and tactile feedback to enable people with visual impairments perform everyday tasks independently. However, a number of issues such as inclusivity concerns, data limitations and technical complexity poses limitations to AI. This article provides an analysis review of AI-driven non-visual accessibility methodologies and solutions for the visually impaired and also explored key challenges faced, and future prospects. By throwing more light on the potential and limitations of non-visual modalities, we aims to guide further research and development in this field.

ChatGPT has transformed the way how education is accessed and made education more interactive and personalised; however, depending on the use, it can be effective or disruptive in education. The use of ChatGPT in education has brought both opportunities and challenges to users and academic institutes. It can make virtually all educational services personalised, automated, efficient and cost-effective; however, users and academic institutes should also be concerned about it using in education due to a number of issues with it, such as misinformation, inaccuracy, bias, cheating, privacy and security; therefore, a careful consideration should be given by them before using it in education. This paper will provide both effective and disruptive aspects of ChatGPT, and elucidate how it can be effective and how it can be disruptive in education. Initially, it will discuss the historical development of chatbots from ELIZA to ChatGPT, and applications of various AI chatbots in education. Afterwards, this paper will elucidate various advantages of ChatGPT in education. Lastly, it will elucidate various challenges of ChatGPT in education. This elucidation will offer a comprehensive and simplified analysis of educational use of ChatGPT.

AI chatbots are progressively becoming more human-like in their creativity and intelligence. Predominantly, AI chatbots based on generative AI models such as Large Language Models (LLMs) and Large Multimodal Model (LMMs) are a significant breakthrough in AI, as they have become more accurate, creative, context-aware, and can better mimic human-like interactions. Generative AI is an umbrella term for all AI technologies used to generate contents, where LLMs can process and generate textual data, and LMMs can process and generate multiple types of data modalities which are also known as multimodal AI models. AI chatbots and multimodal AI models offer several benefits, such as speed, accuracy, creativity, efficiency, and versatility; therefore, they can be used in nearly every sector such as a content creator, proactive assistant, reliable companion, customer assistant, and business operations facilitator. Nonetheless, these AI chatbots and multimodal AI models have some limitations, specially, inaccuracy and hallucination; that can affect and limit their usage for any critical task in any application area. Therefore, human verification and supervision is required for its reliable, safe and guided use. Moreover, their use in any application area also requires appropriate ethical consideration and regulatory compliance, alongside the human verification of generated responses. This paper will present twenty-two applications of AI chatbots and multimodal AI models, which are organised into two categories: personal applications and organisational applications. Subsequently, it will also describe the crucial ethical consideration and regulatory compliance for each application area to emphasise its related ethics, regulations and other risks for humans.

ChatGPT is powered with generative AI and has unlocked the realm of creativity and intelligence to offer virtually all services required by its users. ChatGPT is an AI language model that can understand human language and generate its proficient and coherent responses in human language for the given prompts. ChatGPT has numerous applications such as in education and training, healthcare and medicine, banking and finance, travel and tourism, journalism and media, research and development, customer service and support, and sales and marketing. Despite the fact that ChatGPT is an impressive AI tool and used in almost all application areas, it is not without its flaws and may not always provide accurate or reliable information. Therefore, its use by individuals and businesses requires a quality control over its generated contents to ensure its accuracy and consistency. Moreover, its use in any application area also requires appropriate ethical and regulatory considerations, in addition to the verification of its responses by humans. This paper will explore twenty real-world applications of ChatGPT and its corresponding operations in these application areas. Furthermore, it will also discuss the necessary ethical and regulatory considerations for each application area to highlight its associated ethics, regulations and other risks for users.

Generative AI has transformed the landscape of AI chatbots in a way that they are increasingly becoming human-like in their understanding and working. Generative AI is used for creating new contents, and most modern AI chatbots are designed using two types of generative AI models: Large Language Models (LLMs) and Large Multimodal Model (LMMs). LLMs deal with the text, whereas LMMs deal with more modalities including text, image, audio and video. This continuous enhancement of generative AI models is bringing AI chatbots closer to the human working. Nonetheless, these AI chatbots are not perfect, as they have several limitations, and their users should comprehend before fully relying on these imperfectly perfect AI chatbots. Depending on the type of use of AI chatbots, these limitations may or may not impact them significantly. This paper will organise the limitations of AI chatbots into six main categories: intelligence and understanding related limitations; accuracy and credibility related limitations; ethics and regulations related limitations; accountability, transparency and consistency related limitations; design, coding and training related limitations; and human, machine and vender related limitations. The aim of this paper is to emphasise, organise and analyse numerous limitations of AI chatbots into the proposed categories; accordingly, users should become more cognisant about all these limitations in deciding their suitability or unsuitability for the specific use cases.

ChatGPT is currently one of the most prevalent AI language models, which is constantly astounding its users with each update. It is powered with generative AI and built using a Large Language Model (LLM). ChatGPT demonstrated an exceptional ability to understand natural language whilst generating credible and coherent responses similar to humans. However, as an AI language model, ChatGPT analyses huge amounts of text and generates new text by finding relationships and patterns in the text, without actually understanding the meaning or context of that text, thus lacking human intelligence. Additionally, it has several other limitations related to its development, training, AI model, intelligence, understanding, accuracy, credibility, accountability, transparency, consistency, privacy, security, ethics, and regulations. Each limitation of ChatGPT may not necessarily affect every user; however, it has some consequences for its users depending on the nature of its use and application area. This paper will critically analyse twenty-one limitations of ChatGPT in order to provide a detailed insight into each of its limitation and clarify any hype about it for its balanced study, therefore, users can determine the appropriateness or inappropriateness for its use.

In cellular network technologies like 5G and 6G, achieving high data rates, exceptional reliability, and minimal latency is paramount. The continuous expansion of smart wireless technologies, sensors, and communication systems has led to an exponential increase in data traffic, which increases the demand for enhanced privacy and security solutions that are efficient and reliable as attackers get more aggressive. While numerous authentication and encryption methods have been proposed in the literature to safeguard communication, the landscape of threats and attacks continues to evolve, jeopardizing the security of network entities. In response to this evolving threat landscape, my research introduces an innovative solution to enhance security measures. My approach hinges on dynamic and unpredictable key generation and utilization, setting a new strategy for safeguarding against network breaches. Notably, this work marks the first instance of introducing this mechanism and concept (Key Hopping), promising a more resilient security paradigm. Simulation results demonstrate a high-security performance in most security aspects.

Network slicing (NS) plays a crucial role in 5G networks; it enables the delivery of heterogeneous services and uses cases such as voice communication, video streaming, and Extended Reality (XR), and it also partitions the physical network into several logical networks, known as network slices, to cater the vertical industries via sharing the physical infrastructure. However, infrastructure and functional sharing present security and privacy challenges. In this work, we examine the characteristics of XR traffic across several 5G slicing configurations in the presence of attacks, including ping flood, User Datagram Protocol (UDP) flood, and registration flood attacks. The primary contribution of our research is the analysis of the impact of these different attacks on the XR traffic using different slice configurations. The results indicate the effects of the different attacks on XR traffic in terms of reducing the traffic throughput (Mbps) and the changes in the XR traffic characteristics in the different slice configurations. Also, the isolation of the Virtual Network Functions (VNFs) in the user plane provides better performance in the presence of these attacks.

This paper presents a novel solution to the classical problem of interpolating three-dimensional antenna radiation patterns from two-dimensional, orthogonal pattern slices. We introduce a neural network model that performs this interpolation with high accuracy across a wide range of patterns, including cases where conventional interpolation methods struggle. This model is beneficial for three-dimensional modeling methods, such as ray-tracing, where a full antenna radiation pattern is needed, whereas only principal plane patterns are available from measurements or data sheets. The proposed neural network combines computational efficiency with superior accuracy compared to available alternative methods.

Data availability can limit machine learning model generalizability, while privacy concerns arise when sharing data for collaborative learning. Additionally, limited device capabilities and increased communication power can hinder real-world deployment. Differential-Private-Federated-Learning (DPFL) addresses these challenges by sharing encrypted gradients to enhance privacy and minimize data vulnerability while supporting collaborative model training in privacy-sensitive environments. The integration of DPFL with low-power FP-GAs is beneficial in overcoming edge-device limitations. FP-GAs' parallel processing, reconfigurability, quantization support, and hardware-software co-design, enhance DPFL's efficiency and adaptability. This combined approach optimizes training, ensuring secure and efficient collaboration across applications. Importantly, as ML algorithms evolve, FPGA implementations promote sustainability by reducing energy consumption, extending the hardware lifespan, and reducing electronic waste. Moreover, this framework can be used for emulation on FPGAs to explore and develop novel ASIC architectures for DPFL machine learning accelerators. We propose a novel open-source framework-"DPFL-FPGA-Accel" for carrying out design-space exploration of FPGA-based DPFL, source code is made available on https://github.com/shakeelakram00/DPFL-FPGA-Accel-Framework.git for wider adoption. To the best of our knowledge, this marks the first implementation of DPFL with FPGAs that enables users to design optimized FPGA accelerators for ML tasks implementing privacy-preserving federated learning with adjustable epsilon values. We illustrate the application of this framework in processing healthcare data, designing a DPFLenabled FPGA system for cardiac arrhythmia classification within a typical hospital setting, where electrocardiograms are not shared among hospitals. The FPGA-Accel achieves significant inference speedups, providing up to 605.2x and 107.2x faster performance compared to Arm Cortex-A53 and Intel Core i7 processors, respectively.

In this paper we introduce projection-based precoders for secure transmission in cell-free massive multiple-input multipleoutput (MIMO) systems. Assuming that the channel state information (CSI) of both legitimate users (referred to as Bob) and passive eavesdroppers (Eve) is known, we derive three secure precoder designs using well-known linear precoders, such as maximum ratio transmission (MRT), zero-forcing (ZF) and minimum mean square error (MMSE). The first method projects these linear precoders onto the orthogonal complement space of Eve. The second method processes Bob's channel with the same projection operator before applying linear precoding to this newly formed "secure" channel matrix. Additionally, we propose a novel MMSE-based precoder with added security constraints, designed to reduce energy leaked towards Eve. By reformulating the proposed precoders, we provide several interpretations of the latter, which result in simpler expressions and equivalent projected precoding schemes. Analytical results are supported by simulations in realistic environments, and demonstrate that the proposed methods can completely eliminate Eve's received power. It results in improved secrecy performance compared to their non-secrecy aware counterparts.

We present a novel variant of the transformer encoder architecture designed to enhance predictive ability through sequential bottom-up reasoning with cross-resolution communication. The architecture incorporates cross-scale attention to enable interactions between different hierarchical levels of resolution, auxiliary scale-specific predictors to guide each transformer encoder layer to adopt a distinct resolution in the respective output embedding space, and a variational autoencoder operating on the input embeddings to establish a continuous latent space that facilitates sampling and analysis of how perturbations propagate throughout the network across different levels of resolution.

— The meaningful integration of environmental and neuroimaging information made available by multiple research centers is of significant interest for the study of psychiatric disorders, but is made difficult by having to deal with endogenous and exogenous confounders. To address these open challenges, we developed a pipeline for the extraction of relevant, confounder-free brain-environment latent information with a multimodal convolutional neural network autoencoder (CNN-AE) model. In our study, the model was applied to brain functional connectivity (FC) and environmental data from the multi-site PRONIA cohort, composed of controls and individuals affected with recent-onset depressive (ROD) and psychotic (ROP) disorders. We identified site as a confounder for brain features and diagnosis and sex as outcomes of interest; therefore, we designed the multimodal CNN-AE to integrate brain-environment latent representations, incorporating a strategy for mitigating site effects in the multi-site FC data, facilitating the downstream classification of diagnosis and sex. The site harmonization strategy was designed based on disentanglement representation learning with a site deconfounding block (DB) leveraging a cross-covariance (xcov) penalty. A comparison was performed between the CNN-AE with the site DB placed i) in the brain-related branch, and ii) in the fusion branch, and the CNN-AE without DB, following iii) the state-of-the-art harmonization ComBat model. Our experimental results demonstrate that the CNN-AE with a DB placed directly at the brain-related branch surpassed the state-of-the-art ComBat model in removing site effects and improved the identifiability of sex and diagnosis effects from the fused brain-environment latent space. These findings highlight the potential of embedding the deconfounding processes directly into deep learning (DL) analysis pipelines, offering a more effective approach for addressing confounding variables in neuroimaging studies.

In post-disaster damage assessment, Visual Question Answering (VQA) systems are essential in identifying the severity and scope of damage. However, counting-related tasks, such as determining the number of vehicles and flooded buildings, remain a significant challenge for current deep learning models. To address this issue, we propose DeVANet (DeBERTa Vision Attention Network), a novel architecture aimed at enhancing counting accuracy in VQA for post-disaster scenarios. We leverage DeBERTa for language modeling and introduce an innovative image embedding module, where local-global attention guides Vision Mamba features to achieve precise extraction of both small and large object features. Our fusion mechanism employs self-attention for both text and image features, followed by bi-directional cross-attention and co-attention to enhance multimodal integration. We tackle VQA as both a classification and regression problem by employing separate MLPs for each task: one handling discrete class predictions and the other generating continuous values for counting tasks. A joint loss function, combining weighted cross-entropy and negative binomial loss, ensures optimized performance across both tasks. Extensive experiments on the FloodNet and RescueNet datasets demonstrate that DeVANet achieves significant improvements in counting accuracy and overall VQA performance compared to state-of-the-art works, supported by detailed ablation studies that validate the effectiveness of each component in the architecture.

The operating temperature of lithium-ion battery (LIB) cells significantly influences their degradation behavior. In indirect liquid cooling systems, temperature variations within a Battery Electric Vehicle (BEV) LIB module are inevitable due to the increasing downstream temperature of the cooling medium as it absorbs heat. This leads to reduced temperature differentials between the cooling medium and the LIB cells. As a result, LIB cells located further along the flow path experience higher average temperatures than those at the front. Typically, a maximum average cell temperature difference of 5 K within LIB modules is considered acceptable. However, results from a conventional cooling system indicate that, when fast charging is exclusively used, this can lead to a 15.5 % difference in the total ampere-hours passed before the End-of-Life (EOL) is reached for the front and back LIB cells. To address this issue, a switchable thermal management system for the traction battery is proposed. In this concept, the coolant inlet flow direction is reversed after every fast charge. Compared to conventional systems, this flow direction switch can increase the total ampere-hours passed by 7.2 %, leading to a corresponding 7.2 % increase in vehicle range until the first cell reaches EOL. The concepts are evaluated using a validated 1D electro-thermal-aging model. Additionally, a 3D electro-thermal Computational Fluid Dynamics (CFD) model of a module section, including a wavy-mini-channel cooling design, is employed. Together, these approaches ensure that the 1D model accurately captures the thermal behavior of the LIB module cells.

This study addresses the performance gap between academic benchmarks and real-world applications of Large Language Models (LLMs) in Natural Language to SQL (NL-to-SQL) tasks. We evaluated a diverse range of LLMs across various hosting platforms, using a balanced distribution of 360 questions spanning simple, moderate, and challenging difficulty levels. Our analysis incorporated comprehensive metrics that include execution accuracy, operational costs, and throughput. Results revealed significant platform-dependent performance variations, with models like GPT-4o and Claude-3.5 Sonnet demonstrating superior accuracy (41.44% and 40.94%) but at higher operational costs. In contrast, GPT-4o mini emerged as a cost-effective alternative with 38.56% accuracy. The Mistral family exhibited exceptional throughput, particularly on the Bedrock platform. We observed a consistent decline in performance across all models as query complexity increased, highlighting ongoing challenges in handling intricate database queries. Our findings emphasize the complex interplay between accuracy, cost, and throughput in real-world deployments and underscore the importance of platform-specific optimizations. This study provides valuable insights for researchers and practitioners, offering a nuanced understanding of LLM capabilities in NL-to-SQL tasks and guiding future research directions in this rapidly evolving field. The code is available at https://github.com/petavue/NL2SQL-Benchmark.

The complexity of contemporary language tasks has driven substantial research toward enhancing model architectures capable of processing intricate linguistic patterns with heightened efficiency and contextual precision. Introducing the Dynamic Synaptic Filtering Mechanism (DSFM), a novel approach designed to enable neural pathway modulation, represents a breakthrough in adaptable pathway activation within language models, allowing for dynamic adjustments to input complexity and task-specific demands. Implemented within a transformer-based large language model, DSFM facilitates selective synaptic activation by filtering pathways based on probabilistic assessments, thus offering a level of contextual sensitivity and computational efficiency previously unachievable with traditional static architectures. Comprehensive experiments reveal that DSFM reduces processing latency, optimizes memory allocation, and improves model accuracy, particularly in tasks requiring complex contextual understanding and syntactic variation. Quantitative results indicate that DSFM's probabilistic filtering substantially balances pathway utilization across layers, promoting both adaptive efficiency and computational stability under variable task loads. The findings suggest that DSFM enhances the structural and operational flexibility of language models, achieving a significant reduction in resource demands while maintaining linguistic accuracy, which positions DSFM as a transformative advancement in the pursuit of adaptable, resource-efficient language model architectures.

As a key component of inherent optical properties (IOPs) in ocean color remote sensing, phytoplankton absorption coefficient ($a_{phy}$), especially in hyperspectral, greatly enhances our understanding of phytoplankton community composition (PCC). The recent launches of NASA's hyperspectral missions, such as EMIT and PACE, have created an urgent need for hyperspectral algorithms for studying phytoplankton. Retrieving $a_{phy}$ from ocean color remote sensing in coastal waters has been extremely challenging due to complex optical properties, causing traditional methods to fail, and highlighting the need to improve machine learning approaches, which are hindered by data scarcity, heterogeneity, and noises from data collection. To address this, a novel machine learning framework for hyperspectral retrievals of $a_{phy}$ , based on the mixture-of-experts (MOE), named PhA-MOE, is introduced in this study. Various preprocessing methods for hyperspectral training data are explored, with the combination of robust and logarithmic scalers identified as optimal. In this study, PhA-MOE for $a_{phy}$ prediction is tailored to both past and current hyperspectral missions, including HICO, EMIT, and PACE. Extensive experiments reveal the importance of data pre-processing and improved performance of PhA-MOE in estimating $a_{phy}$ as well as in handling the data heterogeneity.