The deployment of large language models (LLMs) on edge devices and non-server environments presents significant challenges, primarily due to constraints in memory usage, computational power, and inference time. This paper investigates the feasibility of running LLMs across such devices by focusing on optimizing memory usage, employing quantization techniques, and reducing inference time. Specifically, we utilize LLaMA 2 for biomedical text summarization and implement Low-Rank Adaptation (LoRA) quantization to compress the model size to compress the model size and fine-tune it using limited resources. Our study systematically evaluates memory consumption during both training and inference phases, demonstrating substantial reductions through efficient LoRA quantization. Our results indicate that with careful optimization, it is feasible to deploy sophisticated LLMs like LLaMA 2 on low powered devices, thereby broadening the scope of their application in resource-constrained environments.

This research involved designing and building an interactive generative AI application to conduct a comparative analysis of two advanced Large Language Models (LLMs), GPT-4 and Claude 2, using Langsmith evaluation tools. The project was developed to explore the potential of LLMs in facilitating postgraduate course recommendations within a simulated environment at Munster Technological University (MTU). Designed for comparative analysis, the application enables testing of GPT-4 and Claude 2 and can be hosted flexibly on either AWS (Amazon Web Services) or Azure. It utilizes advanced natural language processing and retrieval-augmented generation (RAG) techniques to process proprietary data tailored to postgraduate needs. A key component of this research was the rigorous assessment of the LLMs using the Langsmith evaluation tool against both customized and standard benchmarks. The evaluation focused on metrics such as bias, safety, accuracy, cost, robustness, and latency. Additionally, adaptability covering critical features like language translation and internet access, was independently researched since the Langsmith tool does not evaluate this metric. This ensures a holistic assessment of the LLM’s capabilities.

Online voting is gaining traction in contemporary society to reduce costs and boost voter turnout, allowing individuals to cast their ballots from anywhere with an internet connection. This innovation is cautiously met due to the inherent security risks, where a single vulnerability can lead to widespread vote manipulation. Blockchain technology has emerged as a promising solution to address these concerns and create a trustworthy electoral process. Blockchain offers a decentralized network of nodes that enhances transparency, security, and verifiability. Its distributed ledger and non-repudiation features make it a compelling alternative to traditional electronic voting systems, ensuring the integrity of elections. To further bolster the security of online voting, we propose DemocracyGuard platform on the Ethereum blockchain, which incorporates facial recognition technology to authenticate voters. By leveraging these advancements, DemocracyGuard aims to provide a secure and resilient platform for online voting, paving the way for its broader adoption and revolutionizing the electoral landscape.

Dragon Boat Racing, a popular aquatic folklore team sport, is traditionally held during the Dragon Boat Festival. Inspired by this event, we propose a novel human-based meta-heuristic algorithm called dragon boat optimization (DBO) in this paper. It models the unique behaviors of each crew member on the dragon boat during the race by introducing social psychology mechanisms (social loafing, social incentive). Throughout this process, the focus is on the interaction and collaboration among the crew members, as well as their decision-making in different situations. During each iteration, DBO implements different state updating strategies. By modelling the crew’s behavior and adjusting the state updating strategies, DBO is able to maintain high-performance efficiency. We have tested the DBO algorithm with 29 mathematical optimization problems and 2 structural design problems. The experimental results demonstrate that DBO is competitive with state-of-the-art meta-heuristic algorithms as well as conventional methods.

The idea of a Cyber-Physical-Social System, or CPSS for short, is a relatively new concept that has emerged as a response to the requirement to comprehend the influence that Cyber-Physical Systems (CPS) have on people and vice versa. Conversational assistants (CAs), also called bots, are dedicated to oral or written communication. Over time, the CAs have gradually diversified to today touch various fields such as e-commerce, healthcare, tourism, fashion, travel, and many others sectors. Natural-language understanding (NLU) is fundamental in the Natural Language Processing (NLP) field. Identifying user intents from natural language utterances is a crucial step in conversational systems, and the diversity in user utterances makes intent detection even a challenging problem. Recently, with the emergence of Deep Neural Networks. New State of the Art (SOA) results have been achieved for different NLP tasks. Recurrent Neural networks (RNNs) and recent Transformer architectures are two major players in those improvements. In addition, RNNs have been playing an increasingly important role in sequence modeling in different application areas. On the other hand, Transformer models are new architectures that benefit from the attention mechanism, extensive training datasets, and compute power. First, this review paper presents a comprehensive overview of RNN and Transformer models. Then, a comparative study of the performance of different RNNs and Transformer architectures for the specific task of intent recognition for CAs which is a fundamental task of NLU.

Background: Active Learning with AI-tutoring in Higher Education tackles dropout rates. Objectives: To investigate teaching-learning methodologies preferred by students. ChatGPT-based gamified learning methodology is compared to another active learning methodology and a traditional methodology. Study with Learning Analytics to evaluate alternatives, their implementation, and help students elect the best strategies according to their preferences. Methods: Comparative study of three learning methodologies in a Single-Group counterbalanced with 45 university students. It follows a pretest/post-test approach using AHP and SAM. HRV and GSR used for emotional state estimation. Findings: Criteria related to in-class experiences valued higher than test-related criteria. Chat-GPT integration was well regarded compared to well-established methodologies. Student emotion self-assessment correlated with physiological measures, validating used Learning Analitycs. Conclusions: Proposed model AI-Tutoring classroom integration functions effectively at increasing engagement and avoiding false information. AHP with the physiological measuring allows students to determine preferred learning methodologies, avoiding biases, and acknowledging minority groups.

A document by Leandro A. Passos. Click on the document to view its contents.

This paper presents a new Multi Agent Collaborative Search Algorithm with Adaptive Weights (named MACSAW). MACS is a memetic scheme for multi-objective optimization which contains two kind of actions, the local actions and social actions. The former explore the neighborhood of some virtual agents and the latter push the individual towards the Pareto front. On the base of the latest version of MACS, MACS2.1, we improve the old algorithm from three direction. First, a new kind of utility function is introduced to enhance the convergence. Next, a new social action process which contains more operators and adaptive parameters is embedded in MACSAW. Finally, MACS2.1 lacks the weight vectors adjustment process which leads to diversity losing in some real problems and MACSAW adds it. Further, MACSAW is compared with some state-of-art algorithms and MACS2.1 on some standard benchmarks. It gets competitive results. Two real optimization problems is tackled and the results are analyzed in details.

Recently, Large-scale Language Models (LLMs) such as Chat Generative Pre-trained Transformer (ChatGPT) and Generative Pre-trained Transformer 4 (GPT-4) have demonstrated remarkable performance in the general domain. However, Inadaptability in a particular domain have led to hallucination for these LLMs when responding in specific domain contexts. The issue has attracted widespread attention, existing domain-centered fine-tuning efforts have predominantly focused on sectors like medical, financial, and legal, leaving critical areas such as power energy relatively unexplored. To bridge this gap, this paper introduces a novel power energy chat model called PowerPulse. Built upon the open and efficient foundation language models (LLaMA) architecture, PowerPulse is fine-tuned specifically on Chinese Power Sector Domain Knowledge. This work marks the inaugural application of the LLaMA model in the field of power energy. By leveraging pertinent pre-training data and instruction fine-tuning datasets tailored for the power energy domain, the PowerPulse model showcases exceptional performance in tasks such as text generation, summary extraction, and topic classification. Experimental results validate the efficacy of the PowerPulse model, making significant contributions to the advancement of specialized language models in specific domains.

A document by Said Yasin. Click on the document to view its contents.

The purpose of text classification is to label the text with known labels. In recent years, the method based on graph neural network (GNN) has achieved good results. However, the existing methods based on GNN only regard the text as the set of co-occurring words, without considering the position information of each word in the statement. Meanwhile, this method mainly extracts node features, but neglects the use of edge features between nodes. To solve these problems, a new text classification method, graph convolutional network using positions and edges (PEGCN), is proposed. In the word embedding section, a positional encoding input representation is employed to enable the neural network to learn the relative positional information among words. Meanwhile, the dimension of the adjacency matrix is increased to extract the multi-dimensional edge features. Through experiments on multiple text classification datasets, the proposed method is shown to be superior to the traditional text classification method, and has achieved a maximum improvement of more than 4%.

The analysis of tourist accommodation bookings provides valid information for the management of these establishments. The objective of this work is to analyze the performance of different Machine Learning techniques for the prediction of booking cancellations, as well as to analyze possible patterns in the study data. For this purpose, the following supervised learning methods are used: Multilayer Perceptron Neural Network, Radial Basis Function Neural Network, Decision Tree, Random Forest, AdaBoost and XgBoost, analyzing the performance of these techniques. The dataset used corresponds to the bookings of a resort hotel and a city hotel located in Portugal. As a result, the study compares the classification methods applied and identifies those with better performance, proving that Machine Learning techniques generate reliable forecasts for the management of the tourism industry.

This paper proposes a novel algorithm that accurately predicts market trends and trading entry points for US 30-year-treasury bonds using a hybrid approach of 1-Dimensional Convolutional Neural Network (1DCNN), Long-Short Term Memory (LSTM), and XGBoost algorithms. We compared the performance of various strategies using 1DCNN and LSTM and found that existing state-of-the-art methods based on LSTM have excellent results in market movement prediction tasks, but the effectiveness of 1DCNN and LSTM in terms of trading entry point and market perturbations has not been studied thoroughly. We demonstrate, through experiments that our proposed 1DCNN-BiLSTM-XGBoost algorithm combined with moving averages crossover effectively mitigates noise and market perturbations, leading to high accuracy in spotting trading entry points and trend signals for US 30-year-treasury-bonds. Our experimental study shows that the proposed approach achieves an average of 0.0001% Root Mean Squared Error and 100% R-Square, making it a promising method for predicting the market trends and trading entry points.

The paper proposes an alternative approach to improve the performance of image retrieval. In this work, a framework for image retrieval based on machine learning and semantic retrieval is proposed. In the preprocessing phase, the image is segmented objects by using Graph-cut, and the feature vectors of objects presented in the image and their visual relationships are extracted using R-CNN. The feature vectors, visual relationships, and their symbolic labels are stored in KD-Tree data structures which can be used to predict the label of objects and visual relationships later. To facilitate semantic query, the images use the RDF data model and create an ontology for the symbolic labels annotated. For each query image, after extracting their feature vectors, the KD-Tree is used to classify the objects and predict their relationship. After that, a SPARQL query is built to extract a set of similar images. The SPARQL query consists of triple statements describing the objects and their relationship which were previously predicted. The evaluation of the framework with the MS-COCO dataset and Flickr showed that the precision achieved scores of 0.9218 and 0.9370 respectively.

Purpose: Poly-medicated patients, especially those over 65, have increased. Multiple drug use and inappropriate prescribing increase drug-drug interactions, adverse drug reactions, morbidity, and mortality. This issue was addressed with several CDSS alerts. Health professionals have not followed these systems due to their poor alert quality and incomplete databases. Methods: Recent research shows a growing interest in using Text Mining via NLP to extract drug-drug interactions from unstructured data sources to support clinical prescribing decisions. NLP text mining and machine learning classifier training for drug relation extraction were used in this process. Results: In this context, the proposed solution allows to develop an extraction system for drug-drug interactions from unstructured data sources. The system produces structured information, which can be inserted into a database that contains information acquired from three different data sources. Conclusion: The architecture outlined for the drug-drug interaction extraction system is capable of receiving unstructured text, identifying drug entities sentence by sentence, and determining whether or not there are interactions between them.

Online Social Networks (OSNs) have grown exponentially in the last few years due to their applications in real life like marketing, recommendation systems, and social awareness campaigns. One of the most important research areas in this field is Influence Maximization (IM). IM pertains to finding methods to maximize the spread of information (or influence) across a social network. Previous works in IM have focused on using a pre-defined edge propagation probability or using the Hurst exponent (H) to identify which nodes to be activated. This is calculated on the basis of self-similarity in the time series depicting a user’s (node) past temporal interaction behaviour. In this work, we propose a Time Series Characteristic based Hurst-based Diffusion Model (TSC-HDM). The model calculates Hurst Exponent (H) based on the stationary or non-stationary characteristic of the time series. Furthermore, our model selects a handful of seed nodes and activates every seed node’s inactive successor only if H>0.5 . The process is continued until the activation of successor nodes is not possible. The proposed model was tested on 4 datasets - UC Irvine messages, Email EU-Core, Math Overflow 3, and Linux Kernel mailing list. We have also compared the results against 4 other Influence Maximisation models - Independent Cascade (IC), Weighted Cascade (WC), Trivalency (TV), and Hurst-based Influence Maximisation (HBIM). Our model achieves as much as 590% higher expected influence spread as compared to the other models. Moreover, our model attained 344% better average influence spread than other state-of-the-art models.

The prediction of news popularity is having substantial importance for the digital advertisement community in terms of selecting and engaging users. Traditional approaches are based on empirical data collected through surveys and applied statistical measures to prove a hypothesis. However, predicting news popularity based on statistical measures applied to past data is highly questionable. Therefore, in this paper, we predict news popularity using machine learning classification models and deep residual neural network models. Articles are usually made up of textual content and in many cases, images are also used. Although it is evident that the appropriate amount of textual data is required to extract features and create models, image data is also helpful in gaining useful information. In this paper, we present a novel multimodal online news popularity prediction model based on ensemble learning. This research work acts as a guide for extensive feature engineering, feature extraction, feature selection, and effective modeling to create a robust news popularity Prediction Model. Three kinds of features – meta features, text features, and image features are used to design an influential and robust model. The performance measure Root Mean Squared logarithmic error (RMSLE) is used to validate the outcome of the proposed model. Further, the most important features are sought out for the proposed model to verify the dependence of the model on text and image features.

The repair process of devices is an important part of the business of many original equipment manufacturers. The consumption of spare parts, during the repair process, is driven by the defects found during inspection of the devices, and these parts are a big part of the costs in the repair process. In previous work we proposed a data-driven method for Supply Chain Control Tower solutions to provide support for the automatic check of spare parts consumption in the repair process. In this paper, we continue our investigation of a multi-label classification problem and explore alternatives in the learning-to-rank approach, where we simulate the passage of time using more data while training and comparing hundreds of Machine Learning models to provide an automatic check in the consumption of spare parts. We investigate the effects of different train set sizes, retraining intervals, models and hyper-parameter search using Bayesian Optimization. The results show that we were able to improve the trained models and achieve a higher mean NDCG@20 score of 86% when ranking the expected parts. Focusing on the most recent data, we achieve a NDCG@20 score of 90%, while obtaining a ratio of marked parts of just 4% of the consumed parts for use in alert generation.