March 23 ~ 24, 2024, Sydney, Australia
Qingyue Xiong, Liwei Zhang, and Qiujun Lan, Business School, Hunan University, Changsha-410082, China
Credit risk is the most significant risk faced by credit businesses. Currently, various approaches are widely used in credit evaluation. However, methods based on expert knowledge exhibit obvious subjective cognitive bias, while both statistical and machine learning methods require a substantial amount of historical data. In cases with limited data, the machine-learning effect is poor. Inspired by the structural similarity between neural networks (NN) and the Analytic Hierarchy Process (AHP), we propose a knowledge-augmented dynamic neural network model called KADNN to construct an effective credit evaluation model. This composite architecture will help effectively utilize existing data to alleviate the initial low-data dilemma and can be further utilized for training neural networks. Subsequent data updates can be dynamically incorporated to improve model accuracy. Additionally, this approach improves the comprehensibility and premature convergence issues of the NN model. The proposed approach is validated and evaluated through credit evaluation simulation.
Knowledge Augment, Credit Evaluation, Neural Network, Analytic Hierarchy Process, Machine Learning.
Hasnae Briouya, Asmae Briouya, Ali Choukri, Mohamed Amnai, and Youssef Fakhri, Laboratory of Computer Sciences, Faculty of Sciences Ibn Tofail University, Kenitra, Morocco
This article explores lightweight neural network architectures for mobile devices, with a focus on the widely used MobileNet. We investigate alternative models -ShuffleNet, SqueezeNet, EfficientNet, MnasNet and NASNet Mobile — assessing their balance between simplicity, accuracy, and efficiency. The study aims to guide practitioners in choosing the most suitable architecture based on factors such as computational resources, accuracy requirements, and speed for diverse mobile applications.
MobileNet, Lightweight neural networks, Computational efficiency, Semantic segmentation, Mobile applications.
Dr.Jaekeol Choi, University of New Hampshire, United States of America
Relevance evaluation of a query and a passage is essential in Information Retrieval (IR). Recently, numerous studies have been conducted on tasks related to relevance judgment using Large Language Models (LLMs) such as GPT-4, demonstrating significant improvements. However, the efficacy of LLMs is considerably influenced by the design of the prompt. The purpose of this paper is to identify which specific terms in prompts positively or negatively impact relevance evaluation with LLMs. We employed two types of prompts: those used in previous research and generated automatically by LLMs. By comparing the performance of these prompts, we analyze the influence of specific terms in the prompts. We have observed two main findings. First, performance decreases when a prompt excessively narrows the meaning of relevance. Terms like ‘answer’ or ‘direct’ contribute to this effect. Secondly, the term ‘relevance’ plays a beneficial role. Prompts constructed using the word ‘relevant’ showed improved performance. In conclusion, accurately representing the scope of relevance is essential. There has been no comprehensive study on the essential words in prompts for relevance evaluation. This study will offer valuable insights for IR research using LLMs
chatGPT, GPT-3.5, GPT-4, Information Retrieval, Large Language Models (LLMs), relevance evaluation, prompt engineering, passage ranking.
Alex Xie1, Yu Sun2, 1Chadwick School, 26800 S Academy Dr, Palos Verdes Peninsula, CA 90274, 2California State Polytechnic University, Pomona, CA, 91768, Irvine, CA 92620
This paper presents a comprehensive survey and review of financial prompt patterns, exploring the innovative integration of ChatGPT in finance-related tasks. With the rapid evolution of AI and its increasing application in various sectors, the finance industry stands at the forefront of this technological revolution. Our research delves into the myriad ways in which prompt engineering with ChatGPT can enhance financial analyses, risk assessment, investment strategy formulation, and customer service in the finance sector. We systematically categorize and evaluate a wide array of prompt patterns, drawing insights from real-world applications and theoretical frameworks. This survey not only identifies the current state of prompt engineering in finance but also forecasts future trends, challenges, and opportunities. By providing a detailed examination of various prompt designs and their outcomes, this paper aims to serve as a foundational guide for practitioners and researchers seeking to leverage ChatGPTs capabilities for optimized financial decision-making and innovation. The findings underscore the transformative potential of tailored prompts in elevating the accuracy, efficiency, and scope of financial services and strategies.
Large Language Model, ChatGPT, Finance, Prompt Patterns.
Nidula Elgiriyewithana1, N D Kodikara2, 1Robert Gordon University, Aberdeen, Scotland, 2Informatics Institute of Technology, Colombo, Sri Lanka
In this study, we introduce a novel, parameter-efficient model that employs the attention U-Net architecture for automatic detection and elimination of non-speech vocal sounds, specifically breath sounds, in vocal recordings. This task holds great significance in the realm of sound engineering, despite being under-researched. The manual process to detect and remove thesesounds demands substantial expertise and is highly time-consuming. Previous automated detection and removal methods often lack efficiency and precision. Our proposed model counters thesedrawbacks with its streamlined process and superior accuracy, offering a more efficient approach by leveraging advanced deep learning techniques. A unique dataset, based on Device and Produced Speech (DAPS), was utilized for this purpose. The training phase of the model highlights a log spectrogram and incorporates an early stopping mechanism to circumvent overfitting. Our efforts not only save valuable time for sound engineers but also augment the quality and consistency of audio production. This marks a significant advancement, as demonstrated by its comparative efficiency, requiring only 1.9M parameters and a training time of 3.2 hours - notably less than top-performing models in the field. The model can produce nearly identical outputs as previous models with drastically improved precision, making it an optimal choice.
attention u-net architecture, non-speech vocal sounds, parameter-efficient model, audio quality, deep learning, sound engineering.
Ayoub Nainia1, R´egine Vignes-Lebbe2, Eric Chenin2, Maya Sahraoui3, Hajar Mousannif4, and Jihad Zahir2, 4, 1Sorbonne Universit´e, Mus´eum National d’Histoire Naturelle, Institut de Syst´ematique Evolution Biodiversit´e, ISYEB, Paris, France, 2UMMISCO, IRD France Nord, Bondy, 3Mus´eum National d’Histoire Naturelle, Paris, France, 4LISI Laboratory, Cadi Ayyad University, Marrakesh, Morocco
Abstract. This research investigates the untapped wealth of centuries-old French botanical literature, particularly focused on floras, which are comprehensive guides detailing plant species in specific regions. Despite their significance, this literature remains largely unexplored in the context of AI integration. Our objective is to bridge this gap by constructing a specialized botanical French dataset sourced from the flora of New Caledonia. We propose a transformer-based Named Entity Recognition pipeline, leveraging distant supervision and CamemBERT, for the automated extraction and structuring of botanical information. The results demonstrate the efficacy of our approach, showcasing high recall, precision, and F-score across all models. This work contributes to the exploration of valuable botanical literature by underscoring the capability of AI models in automating information extraction from complex and diverse texts.
Information Extraction, Natural Language Processing, Named Entity Recognition, Biodiversity Literature.
Sarah Alnefaie, Eric Atwell and Mohammad Ammar Alsalka, University of leeds, United Kingdom
One of the challenges in natural text processing concerns answering questions, and studies have tended to use large language models for this task. However, these models often generate misleading and incorrect answers. Answering religious questions is a delicate and sensitive topic. Therefore, the contribution of this paper is to use the retrieval-augmented generation technique to improve the performance of large language modules in Arabic. In this experiment, we compared the answers of standard GPT-4 and GPT-4 using the retrieval-augmented generation technique for the same questions on the Quranic Question–Answer Pair dataset. The results showed that retrieval-augmented generation technique improved the answers to different types of questions. It also improved the text parts of answers, reaching a 0.266 exact match (EM) and 0.772 F1. In addition, it improved the parts of answers regarding Quranic verses, achieving 0.333 F1@1, 0.25 EM, 0.53 F1, and 0.266 partial average precision.
Retrieval-augmented generation technique, GPT-4, large language model, Quranic Question Answering model.
Mashaal A. M. Saif, and Nazimah Hussin, Accounting and Finance Department, Azman Hashim International Business School, Universiti Teknologi Malaysia, Kuala Lumpur 54100, Malaysia
This study employs an Artificial Neural Network (ANN) to predict the importance of various factors—convenience, economic efficiency, functional risk, security risk, critical mass, number of services, trust, environmental concern, and perceived value—in collectively determining the intention to adopt digital-only banks. The analysis involves 403 respondents in Malaysia, utilising both exploratory factor analysis and the ANN method. The results from the ANN highlight "environmental concern" as the most influential factor shaping individuals intention to adopt digital-only banks. Additionally, "trust," "perceived value," and "convenience" emerged as crucial factors in predicting adoption intention. These findings not only provide valuable insights for fintech companies and banks aiming to attract new customers or enter new markets but also contribute to expanding knowledge in the field, particularly in the realm of non-linear methods such as ANN. This study enhances the understanding of the evolving landscape in digital-only banking.
Artificial Neural Network (ANN); Adoption Intention; Digital-only banks; Perceived Value; Environmental Concern.
Mehmet Altug Akgul, Hakan Guvez
The emergence of microservice architecture and the concept of Chaos Engineering, which was researched to understand how the system will behave in case of possible interruptions and service problems that may occur when software services trying to do the same job in different locations talk to each other, play an important role in determining the robustness of systems exposed to intense network traffic. In this article, the concept of Chaos Engineering, which has emerged in recent years and has started to be used especially in checking the reliability of distributed systems, is mentioned. As a result, it is emphasized that Chaos Engineering can play an important role in increasing the reliability of distributed systems running on the cloud.
cloud, chaos engineering, distributed systems, microservices, infrastructure.
Subhadip, Kumar, Western Governors University
The adoption of Automatic Train Operation (ATO) has revolutionized urban metro systems, enhancing safety, efficiency, and passenger experience. However, the application of ATO in freight railways remains underexplored. This research investigates the challenges specific to freight railways, including extended braking distances and obstacles on rail tracks. Leveraging recent advancements in Artificial Intelligence (AI), Computer Vision (CV), and Machine Learning (ML), we propose innovative solutions to address these challenges. Our study focuses on the strategic integration of AI, CV, and ML technologies within the context of freight railways. By analyzing past ATO studies and examining the impact of advanced vision sensors (such as advanced signaling, cameras, lidars, drones, and radars), we aim to pave the way for efficient, safe, and automated freight transportation. This research contributes to the ongoing transformation of the rail industry, emphasizing the need for collaboration between technology, infrastructure, and operational practices.
IoT, Artificial Intelligence, ATO.
Gyu Bae1, Byung Jun Kwon2, 1University of California, Berkeley, Seoul, Republic of Korea, 2Cornerstone Collegiate Academy of Seoul, Seoul, Republic of Korea
We introduce a novel string-based soft robotic gripper, capable of gentle and precise manipulation of delicate and small objects. The primary challenge lies in delicately handling objects of varying shapes and sizes without causing damage, a prevalent issue with traditional hard grippers. Our solution employs a string-based design, where the strings envelop the object, providing secure and damage-free grip. A motor torque sensor is utilized to optimally control the gripping pressure, preventing over-gripping and ensuring object integrity. We have successfully tested our soft gripper on delicate items ranging from grains of rice to strawberries. This innovative approach presents a solution to two pressing problems: the labor-intensive and potentially damaging manual picking of delicate items, and the inability of conventional robotic grippers to handle small and delicate objects effectively. Furthermore, the versatility and delicate handling of our soft gripper opens up potential applications in various fields such as agriculture, healthcare, and manufacturing. The design and performance of the soft gripper, its advantages over traditional grippers, and its potential applications are explored in this paper. We demonstrate the efficacy of our design through experimental results, confirming the robustness and adaptability of the string-based soft gripper.
Soft Robotic Gripper, String-Based Design, Delicate Object Manipulation, Small Object Handling, Motor Torque Control.
This study focuses on applying explainable artificial intelligence approaches to solve the challenge of attack detection in Internet of Things networks. The paper emphasizes how the proposed intrusion detection platform might enhance the model ability of explaining and comprehending attacks that might take place. The study suggests theoretical and practical advancements that directly affect the intrusion detection in Internet of Things networks. To comprehend how the system operates and how attacks happen, this paper describes the architecture and structure of the introduced explainable model. The targeted contributions focus on identifying possible attacks and interpreting taken decisions. The findings of testing on the UNSW-NB15 dataset show a better performance of RNN-SHAP model when compared to other algorithms (GRU, SVM, and RL), in terms of numerous metrics such as accuracy (98.98%), precision (100%), recall (93.44%), and F1-score (95.63%).
Attack detection, explainable artificial intelligence (XAI), deep learning (DL), Recurrent neural networks (RNN), SHapley Additive explanation (SHAP), Distributed Denial of Service (DDoS)