January 27 ~ 28, 2024, Copenhagen, Denmark
Sinan Gultekin, Achille Globo, Andrea Zugarini,Marco Ernandes, and Leonardo Rigutini, Department of Hybrid Linguistic Technologies.expert.ai, Siena, Italy
Most Machine Learning research evaluates the best solutions in terms of performance. However, in the race for the best performing model, many important aspects are often overlooked when, on the contrary, they should be carefully considered. In fact, sometimes the gaps in performance between different approaches are neglectable, whereas factors such as production costs, energy consumption, and carbon footprint must take into consideration. Large Language Models (LLMs) are extensively adopted to address NLP problems in academia and industry. In this work, we present a detailed quantitative comparison of LLM and traditional approaches (e.g. SVM) on the LexGLUE benchmark, which takes into account both performance (standard indices) and alternative metrics such as timing, power consumption and cost, in a word: the carbon-footprint. In our analysis, we considered the prototyping phase (model selection by training-validation-test iterations) and in-production phases separately, since they follow different implementation procedures and also require different resources. The results indicate that very often, the simplest algorithms achieve performance very close to that of large LLMs but with very low power consumption and lower resource demands. The results obtained could suggest companies to include additional evaluations in the choice of Machine Learning (ML) solutions.
NLP, text mining, green AI, green NLP, carbon footprint, energy consumption, evaluation.
Mahdi Shafiei and Kirsten Wahlstrom, STEM department of University of South Australia, Adelaide, Australia
Authorship attribution is a key process in text mining with numerous applications, particularly in the field of social networks. It is defined as identifying an author based on the extracted patterns from the same author in the past. In this paper, we continue the trend of narrowing the scope of the features, which has been trending from sentences to characters in the literature and focus on the binary representation of the text. We develop a method to extract binary n-gram features from the text and then train a classification model to predict the author based on the provided set of the features. Our approach can use either short or medium-size texts or their mixture without separating them. We evaluate our approach using a number of various data sets including comments and reviews on social platforms.
Authorship attribution, text feature extraction, n-gram.
Kamal Sarkar1 and Sohini Roy Chowdhury2, 1Jadavpur University, 188, Raja S.C. Mallick Rd, Kolkata-700032, West Bengal, India, 2RCC Institute Of Information Technology, Canal South Rd, Beleghata, Kolkata 700015, West Bengal, India
The process of creating a single summary from a group of related text documents obtained from many sources is known as multi-document summarization. The efficacy of a multidocument summarization system is heavily reliant upon the sentence similarity metric employed to eliminate redundant sentences from the summary, given that the documents may contain redundant information. The sentence similarity measure is also crucial for a graph-based multi-document summarization, where the presence of an edge between two phrases is decided by how similar the two sentences are to one another. To enhance multi-document summarization performance, this study provides a new method for defining a hybrid sentence similarity measure combining a lexical similarity measure and a BERT-based semantic similarity measure. Tests conducted on the benchmark datasets demonstrate how well the proposed hybrid sentence similarity metric is effective for enhancing multi-document summarization performance.
Extractive Summarization. Multi-Document Text Summarization. BERT. Hybrid Similarity measure. Semantic Similarity similarity, Lexical similarity.
Ilias Boulbarj1 Bouklouze Abdelaziz1 Yousra El Alami2 Douzi Samira3,4 Douzi Hassan1, 1IRF-SIC Laboratory ,Ibn Zohr Universitys -Agadir, Morocco, 2Laboratory of Pharmacology and Toxicology, Faculty of Medicine and Pharmacy, Mohammed V University, Morocco, 3Faculty of Sciences, IPSS Laboratory, Mohammed V University, Rabat, Morocco, 4Faculty of Medicine and Pharmacy, Mohammed V University, Rabat, Morocco
Honey, an organic and highly esteemed dietary substance, is vulnerable to adulteration, hence presenting significant implications for both public health and economic welfare. The conventional techniques employed for the identification of honey adulteration are characterized by prolonged duration and frequently exhibit limited sensitivity. This paper introduces a novel methodology to tackle the aforementioned problem by utilizing Convolutional Neural Networks (CNNs) to classify honey samples using thermal pictures. Thermal imaging technology possesses a distinct edge in the identification of adulterants, as it has the capability to unveil temperature discrepancies within honey samples resulting from variances in sugar composition, moisture levels, and additional adulterating agents. In order to develop a reliable and precise approach for categorizing honey, we gathered a comprehensive dataset consisting of thermal pictures of genuine and contaminated honey samples. Multiple convolutional neural network (CNN) architectures were trained and fine-tuned using this dataset, the findings of our study showcase the capacity of thermal image analysis in conjunction with convolutional neural networks (CNNs) as an effective instrument for promptly and accurately identifying instances of honey adulteration. The methodology presents a potentially advantageous pathway for implementing quality control measures within the honey business, thereby guaranteeing the authenticity and safety of this valuable organic commodity.
Honey adulteration, Inception Net, CNN, Thermal imaging, Quality control, human health.
Theo Zangato, Osmani, Sorbonne Paris Nord University - LIPN-UMR CNRS 7030, France
Amidst increasing energy demands and growing environmental concerns, the promo- tion of sustainable and energy-efficient practices has become imperative. This paper introduces a reinforcement learning-based technique for optimizing energy consumption and its associated costs, with a focus on energy management systems. A three-step approach for the efficient man- agement of charging cycles in energy storage units within buildings is presented combining RL with prior knowledge. A unique strategy is adopted: clustering building load curves to discern typical energy consumption patterns, embedding domain knowledge into the learning algorithm to refine the agent’s action space and predicting of future observations to make real-time decisions. We showcase the effectiveness of our method using real-world data. It enables controlled explo- ration and efficient training of Energy Management System (EMS) agents. When compared to the benchmark, our model reduces energy costs by up to 15%, cutting down consumption during peak periods, and demonstrating adaptability across various building consumption profiles.
Reinforcement Learning, Energy Management Systems, Time-Series, Clustering
Ram Sivaraman1, Joe Xiao2, 1Liberal Arts and Science Academy, Austin, Texas, USA, 2Optum/UnitedHealthCare, Minneapolis, Minnesota, USA
An electrocardiogram (ECG) is a common method used for diagnosis of heart diseases. ECG is not sufficient to detect heart abnormalities early. Heart sound monitoring or phonocardiogram (PCG) is a non-invasive assessment that can be performed during routine exams. PCG can provide valuable details for both heart disorder diagnosis as well as any perioperative cardiac monitoring. Further, heart murmurs are abnormal signals generated by turbulent blood flow in the heart and are closely associated with specific heart diseases.This paper presents a new machine learning-based heart sounds evaluation for murmurs with high accuracy. A random forest classifier is built using the statistical moments of the coefficients extracted from the heart sounds. The classifier can predict the location of the heart sounds with over 90% accuracy. The random forest classifier has a murmur detection accuracy of over 70% for test dataset and detects with over 98% accuracy for the full dataset.
Random Forest Network, Phonocardiogram, Heart Murmur, Sound Features
Xabier Echeberria-Barrio, Mikel Gorricho, Selene Valencia, and Francesco Zola, Vicomtech Foundation, Basque Research and Technology Alliance (BRTA); Paseo Mikeletegi, 57, Donostia 20009, Spain
The usage of Artificial Intelligence (AI) systems has increased exponentially, thanks to their ability to reduce the amount of data to be analyzed, the user efforts and preserving a high rate of accuracy. However, introducing this new element in the loop has converted them into attacked points that can compromise the reliability of the systems. This new scenario has raised crucial challenges regarding the reliability and trustworthiness of the AI models, as well as about the uncertainties in their response decisions, becoming even more crucial when applied in critical domains such as healthcare, chemical, electrical plants, etc. To contain these issues, in this paper, we present NeuralSentinel (NS), a tool able to validate the reliability and trustworthiness of AI models. This tool combines attack and defence strategies and explainability concepts to stress an AI model and help non-expert staff increase their confidence in this new system by understanding the model decisions. NS provide a simple and easy-to-use interface for helping humans in the loop dealing with all the needed information. This tool was deployed and used in a Hackathon event to evaluate the reliability of a skin cancer image detector. During the event, experts and non-experts attacked and defended the detector, learning which factors were the most important for model misclassification and which techniques were the most efficient. The event was also used to detect NS’s limitations and gather feedback for further improvements.
Adversarial Attack, Defence Strategy, Trustworthiness AI, Explainability, Human-AI Teaming.
Ricardo de Deijn, Aishwarya Batra, Brandon Koch, Hema Makkena and Naseef Mansoor, Department of Computer Information Science, Minnesota State University, Mankato, United States of America
The growth of generative models has increased the ability of image processing and provides numerous industries the technology to produce realistic image transformations. However, with the field being newly established there are new evaluation metrics that can further this research. Previous research has shown the FID to be an effective metric when testing these image-to-image GANs in real-world applications. SID, a founded metric in 2023, expands on FID by allowing unsigned distances. This paper uses data that consists of building façades, cityscapes, and maps within Pix2Pix and CycleGAN models. After training these models, we evaluate the models on both the FID and SID metrics. FID is a standalone metric within image-to-image GAN and is commonly used to assess GAN performances. Our findings indicate that the creation of SID incorporates a new efficient and effective metric to complement, or even exceed the ability shown using the FID for the translation GANs.
Signed Inception Distance, Fréchet Inception Distance, Generative Adversarial Networks, Supervised Image-to-Image Translation.
Wouter Knibbe, Informatics Institute, University of Amsterdam, Amsterdam, The Netherlands
Games and agents are the two main components of Reinforcement Learning (RL). While the literature on agents is extensively detailed, little knowledge is shared about how the games they play should be designed. Through a data-science-oriented, integrative literature review, a body of RL literature is analyzed. It was found that no papers reference a method for RL environment design, game design, or a method that serves a similar purpose. In an attempt to create such a method, an evidence-based guide is synthesized which summarizes the best practices in RL environment design. The results show a methodological difference between RL environments in general and those that allow researchers to study and solve real-world systems. The latter we define as AI games with a purpose, or AI games for short. While the results of this study can be used to inform design decisions for future AI games, it is also based on a body of literature that is fundamentally lacking and largely unreproducible. This study should be seen as a zero measurement from which the literature can start to catch up to advancements in the field and develop its methodology of modern RL environment design.
Artificial Intelligence, Reinforcement Learning, Methodology, Literature Review, Games, Real-world AI.
Negin Sokhandan, Mohammad Mahmudul Alam, Emmett Goodman, Amazon, United States of America
In Virtual Product Placement (VPP) applications, the discrete integration of specific brand products into images or videos has emerged as a challenging yet important task. This paper introduces a novel three-stage fully-automated VPP system. In the first stage, a language-guided image segmentation model identifies optimal regions within images for product inpainting. In the second stage, Stable Diffusion (SD), fine-tuned with a few example product images, is used to inpaint the product into the previously identified candidate regions. The final stage introduces an ’Alignment Module’, which is designed to effectively sieve out low-quality images. Comprehensive experiments demonstrate that the Alignment Module ensures the presence of the intended product in every generated image, and enhances the average quality of images by 35%. The results presented in this paper demonstrate the effectiveness of the proposed VPP system, which holds significant potential for transforming the landscape of virtual advertising and marketing strategies.
Wouter Knibbe, Informatics Institute, University of Amsterdam, Amsterdam, The Netherlands
Games and agents are the two main components of Reinforcement Learning (RL). While the literature on agents is extensively detailed, little knowledge is shared about how the games they play should be designed. Through a data-science-oriented, integrative literature review, a body of RL literature is analyzed. It was found that no papers reference a method for RL environment design, game design, or a method that serves a similar purpose. In an attempt to create such a method, an evidence-based guide is synthesized which summarizes the best practices in RL environment design. The results show a methodological difference between RL environments in general and those that allow researchers to study and solve real-world systems. The latter we define as AI games with a purpose, or AI games for short. While the results of this study can be used to inform design decisions for future AI games, it is also based on a body of literature that is fundamentally lacking and largely unreproducible. This study should be seen as a zero measurement from which the literature can start to catch up to advancements in the field and develop its methodology of modern RL environment design.
Artificial Intelligence, Reinforcement Learning, Methodology, Literature Review, Games, Real-world AI.
Suliman Alnutefy and Ali Alsuwayh, School of Technology and Innovation, Marymount University – Ballston Center, United States
This research focuses on Unsupervised Anomaly Detection using the "ambient_temperature_system_failure.csv" dataset from Numenta Anomaly Benchmark (NAB). The dataset contains time-series temperature readings from an industrial machines sensor. The aim is to detect anomalies indicating system failures or aberrant behavior without labeled data. Various algorithms, such as K-means, Gaussian/Elliptic Envelopes, Markov Chain, Isolation Forest, One-Class SVM, and RNNs, are applied to analyze the temperature data. These algorithms are chosen for their ability to identify significant deviations in unlabeled datasets. The study explores how these techniques enhance anomaly understanding in time series data, relevant in manufacturing, healthcare, and finance. This researchs novelty lies in employing unsupervised learning techniques on a real-world dataset and understanding their adaptability in anomaly detection. The results are expected to contribute valuable insights to the field, showcasing the practicality and effectiveness of these algorithms across various scenarios.
Unsupervised Anomaly Detection, Time Series Data, Numenta Anomaly Benchmark, Industrial Machine Sensor Data, Algorithm Analysis, Machine Learning.
FupingRen1, Jian Chena2, DefuZhanga2, 1Shenzhen Comtech Technology Co. Ltd, Shenzhen 518063,China, 2School of informatics, Xiamen University,Xiamen,361005, China
Understanding complex policy documents can be challenging, highlighting the need for intelligent interpretation of Chinese policies. To enhance Chinese text summarization, this study utilized the mT5 model as the core framework and initial weights. Additionally, it reduced model size through parameter clipping, employed the Gap Sentence Generation (GSG) method as an unsupervised technique, and enhanced the Chinese tokenizer. After training on a meticulously processed 30GB Chinese training corpus, the study developed the enhanced mT5-GSG model. When fine-tuning on Chinese policy texts, it adopted the "Dropout Twice" approach and ingeniously merged the probability distribution of the two dropouts using the Wasserstein distance. Experimental results indicate that the proposed model achieved Rouge-1, Rouge-2, and Rouge-L scores of 56.13%, 45.76%, and 56.41% respectively on the Chinese policy text summarization dataset.
Natural Language Processing, Text Summarization, Transformer model.
Jason S. Chu1 and Sindhu Ghanta2, 1Monta Vista High School, Cupertino, CA, USA, 2AIClub, Mountain View, CA, USA
Exploring the area of multimodal sentiment analysis, this paper addresses the growing significance of this field, driven by the exponential rise in multimodal data across platforms like YouTube. Traditional sentiment analysis, primarily focused on textual data, often overlooks the complexities and nuances of human emotions conveyed through audio and visual cues. Addressing this gap, our study explores a comprehensive approach that integrates data from text, audio, and images, applying state-of-the-art machine learning and deep learning techniques tailored to each modality. Our methodology is tested on the CMU-MOSEI dataset, a multimodal collection from YouTube, offering a diverse range of human sentiments. Our research highlights the limitations of conventional text-based sentiment analysis, especially in the context of the intricate expressions of sentiment that multimodal data encapsulates. By fusing audio and visual information with textual analysis, we aim to capture a more complete spectrum of human emotions. Our experimental results demonstrate notable improvements in precision, recall and accuracy for emotion prediction, validating the efficacy of our multimodal approach over single-modality methods. This study not only contributes to the ongoing advancements in sentiment analysis but also underscores the potential of multimodal approaches in providing more accurate and nuanced interpretations of human emotions.
Transformers, multi-modal, sentiment analysis.
Afrooz Rahmati, Afra Mashhadi, and Geethapriya Thamilarasu, Computing and Software Systems, University of Washington Bothell,Bothell, WA USA
The Internet of Things (IoT) has emerged as the next big technologi-cal revolution in recent years with the potential to transform every sphere of human life. As devices, applications, and communication networks become increasingly con- nected and integrated, security and privacy concerns in IoT are growing at an alarm- ing rate as well. While existing research has largely focused on centralized systems to detect security attacks, these systems do not scale well with the rapid growth of IoT devices and pose a single-point of failure risk. Furthermore, since data is exten- sively dispersed across huge networks of connected devices, decentralized computing is critical. Federated learning (FL) systems in the recent times has gained popularity as the distributed machine learning model that enables IoT edge devices to collabo- ratively train models in a decentralized manner while ensuring that data on a user’s device stays private without the contents or details of that data ever leaving that device. In this paper, we propose a federated learning based intrusion detection sys- tem using LSTM Autoencoder. The proposed technique allows IoT devices to train a global model without revealing their private data, enabling the training model to grow in size while protecting each participants local data. We conduct extensive experiments using the BoT-IoT data set and demonstrate that our solution can not only effectively improve IoT security against unknown attacks but also ensure users data privacy.
Internet of Things, security, Intrusion Detection system, Federatedlearning, Deep embedded clustering.
Hemendra Vyas, Computer Science and Engineering, NYU Tandon School of Engineering, New York, NY
Data is growing enormously across all industries, banking and financial institutions are no exception. Financial organizations are increasingly interested in effectively managing and using day to day data to make business decisions and complying with new and existing regulations. There are general regulatory requirements for data retention of up to 7 years which makes the overall data management process challenging. To overcome this challenge banks and financial institutes rely on regular data backups of individual applications. With new regulations such as Fundamental Review of Trading Books being implemented in 2023-24, which impact multiple areas of bank, there is an immediate need for a centralized database to handle big data. In this paper author proposes a big data platform for a typical investment bank which can unify the data needs of Trading, Market Risk, Credit Risk, Counterparty Risk, Enterprise Risk Management and Model Risk Management and help with regulatory compliance.
Data management, Big Data, Banking Regulations, Market Risk and Trading Systems, Compliance, Models.
Vanshaj Gupta1, Jaydeep Patela1, Safa Shubbara1, and Kambiz Ghazinoura2, 1Department of Computer Science, Kent State University, Kent, OH USA, 2Department of Cybersecurity, State University of New York at Canton, Canton, NY, USA
In a time when social media information is a valuable resource for gaining insights, the COVID-19 pandemic has released a flood of public sentiment, abundant with unstructured text data. This paper introduces CovBERT, a novel adaptation of the BERT model, specifically honed for the nuanced analysis of COVID-19related discourse on X (formerly Twitter). CovBERT stands out by incorporating a bespoke vocabulary, meticulously curated from pandemic-centric tweets, resulting in a remarkable leap in sentiment analysis accuracy—from the baseline 72% to an impressive 78.64%. This paper not only presents a detailed comparison of CovBERT with the standard BERT model but also juxtaposes it against traditional machine learning approaches, showcasing its superior proficiency in decoding complex emotional undercurrents in social media data. Furthermore, the integration of geolocation analysis pipeline adds another layer of depth, offering a panoramic view of global sentiment trends.
BERT, CovBERT, COVID-19, Sentiment Analysis, X (Twitter) Data Analysis, Natural Language Processing, Machine Learning, Geolocation Analysis, Social Media Analytics, Data Mining.