ABSTRACT. Natural Language Understanding is a core task when building conversational agents, fulfilling the objectives of understanding the user’s goal and detecting any valuable information regarding it. NLU implies Intent Detection and Slot Filling, to semantically parse the user’s utterance. One caveat when training a Deep Learning model for a domain specific NLU is the lack of specific datasets, which leads to poorly performing models. To overcome this, we experiment with fine-tuning BERT to jointly detect the user’s intent and the related slots, using a custom-generated dataset built around a organization specific knowledge base. Our results show that well-constructed datasets lead to high detection performances and the resulting model has the potential to enhance a future task-oriented dialogue system.

ABSTRACT. We tackle the task of enriching ontologies by automatically translating natural language (NL) into Description Logic (DL). Since Large Language Models (LLMs) are the best tools for translations, we fine-tuned a GPT-3 model to convert NL into OWL Functional Syntax. For fune-tuning, we designed pairs of sentences in NL and the corresponding translations. This training pairs cover various aspects from ontology engineering: instances, class subsumption, domain and range of relations, object properties relationships, disjoint classes, complements, or cardinality restrictions.The resulted axioms are used to enrich an ontology, in a human supervised manner. The developed tool is publicly provided as a Protégé plugin.

ABSTRACT. Age-Related Macular Degeneration (AMD) represents one of the leading causes of blindness for individuals over 65. Optical Coherence Tomography represents a noninvasive examination that not only can help clinicians diagnose multiple retinal abnormalities, AMD included, but also monitor the progression of the disease. There are still unmet needs in terms of personalized treatment for patients suffering from AMD, being a disease that displays individual diversity in terms of its progression and outcomes. We propose a method that will use deep learning methodologies to analyze patients’ disease progression and possible outcomes only based on the OCT scans that are taken during the initial first two examinations. This paper will propose an architecture that will help medical professionals with their involvement in the administration of antiVEGF injections, which are the standard treatment for advanced neovascular AMD. The above-mentioned architecture is based on features extracted from B-scans of an OCT volume and through transfer learning practices will predict the total amount of injections required for an individual who is under treatment as well as next-visit injection administration.

ABSTRACT. This paper presents a deep learning-based model for multi-temporal urban development prediction using satellite imagery. The model is designed to forecast the evolution of urban areas in both the past and the future, enabling informed decision-making in urban planning. The model achieves enhanced performance through iterative improvements and adaptations in hyperparameters, augmentation techniques, loss functions, and model structure. Fine-tuning and activation function adjustments further optimize the model's predictive capabilities. Evaluation of diverse datasets showcases the model's robustness and applicability. The predictions provide valuable insights for sustainable urban growth and support evidence-based decision-making in urban development. The key findings highlight the impact of various improvements on the model's performance. Notably, introducing the time skip as an Embedding layer proved to be a valuable choice, enabling the model to capture temporal dependencies more effectively. Additionally, shifting our focus to the pixel-level differences between the target and input images provided the model with a more informative learning signal, leading to improved predictions. These enhancements, coupled with hyperparameter optimization, suitable augmentations, and adjustments in the loss calculation, collectively contributed to significant advancements in the model's overall performance.

ABSTRACT. Social networks are full of news, opinions or research studies, however, some of the information provided might not be completely true. Furthermore, some people can be easily influenced to take over other people’s beliefs without researching the veracity of the idea. This leads to the existence of various conspiracy theories. In this era of the abundant information existent about any subject, it is very important to be correctly informed. In this work, we investigate several Facebook posts which belong to various conspiracy theories. Our goal is to develop a system that can automatically summarize the information from these posts more objectively. For this, we use a Kernel Ridge Regression(KRR) model which transforms a subjective post into another text description that is more objective. We collected a data corpus from Facebook which contains written posts to misinform the population about various conspiracy theories, such as population control through chips in vaccines, controlling the virus through 5G, the virus being created in a lab by Americans / Chinese, the face masks are dangerous, the government wants to impose a dictatorship, the reporting of cases is wrong, the COVID-19 pandemic, and so on. The posts were generally collected from the same Facebook accounts, which belong to some influencers and public figures. We compared our proposed method for summary generation with a deep learning approach. The experimental results show that our proposed method for generating text summaries performs better than applied deep-learning approaches. Furthermore, the proposed model generated similar words to the original posts.

ABSTRACT. Automatic semantic change methods try to identify the changes that appear over time in the meaning of words by analyzing their usage in diachronic corpora. In this paper, we analyze different strategies to create static and contextual word embedding models, i.e., Word2Vec and ELMo, on real-world English and Romanian datasets. To test our pipeline and determine the performance of our models, we first evaluate both word embedding models on an English dataset (SEMEVAL-CCOHA). Afterward, we focus our experiments on a Romanian dataset, and we underline different aspects of semantic changes in this low-resource language, such as meaning acquisition and loss. The experimental results show that, depending on the corpus, the most important factors to consider are the choice of model and the distance to calculate a score for detecting semantic change.

ABSTRACT. This study shows the generation process and the subsequent study of the representation space obtained by extracting GLCM texture features from computer-aided tomography (CT) scans of pulmonary nodules (PN). For this, data from 92 patients from the Germans Trias i Pujol University Hospital were used. The workflow focuses on feature extraction using Pyradiomics and the VGG16 Convolutional Neural Network (CNN). The aim of the study is to assess whether the data obtained have a positive impact on the diagnosis of lung cancer (LC). To design a machine learning (ML) model training method that allows generalization, we train SVM and neural network (NN) models, evaluating diagnosis performance using metrics defined at slice and nodule level.

ABSTRACT. Today, the Transformer Neural Network (TNN) architecture provides top results in many data processing applications (text, voice, image, and video), outperforming the more traditional convolutional or recurrent Deep Neural Network (DNN) models. To improve AI-based automated diagnosis in the field of medicine, this research seeks to explore TNN capabilities for the characterization of focal liver lesion (FLL) using contrast-enhanced ultrasound (CEUS). Firstly, our study reviewed TNN architectures used in image classification tasks, and then it aimed at the identification of a suitable TNN variant for the above-mentioned topic. This later aim is justified by the fact that, in a typical case, a TNN works much better when it has a large amount of data available for the training process. Unfortunately, this is not the case for most available CEUS datasets, including the one considered in this paper. We compared our proposal with other solutions based on machine learning reported in the literature and found that it provides comparable accuracy. Moreover, this is done by classifying a higher number of FLL types than most previous CEUS Computer Aided Diagnosis (CAD) systems.

ABSTRACT. Background. One of the most prevalent autoimmune diseases and the main contributor to hypothyroidism in regions with sufficient iodine levels is Hashimoto's thyroiditis. A theory that arose in recent years suggested that thyroid autoimmunity might be linked to low-grade chronic inflammation, which may cause cardiovascular comorbidities in the future, independent of thyroid function. Therefore, it is crucial to identify Hashimoto's thyroiditis early on and do thyroid function tests.

Methods. We gathered 129 volunteers, 104 of whom had been diagnosed with Hashimoto's thyroiditis, and 25 controls that did not have this disease. Secondly, we gathered 12 factors and examined their significant differences between controls and Hashimoto's thyroiditis patients. The clinical factors analyzed were age, family history of autoimmune thyroid disease, personal history of breast cancer, surgically induced menopause, diabetes mellitus type 2, and polycystic ovary syndrome. The following paraclinical parameters were examined: hypertriglyceridemia, anemia, hemoglobin and hematocrite levels. hypercholesterolemia abnormal liver function tests, hyperuricemia, and fasting hyperglycemia. For classification and regression, we assessed the following machine learning models: Decision Tree, K-Nearest Neighbors, Extreme Gradient Boost, Support Vector Machine, as well as Artificial Neural Network and Deep Neural Network.

Results. Extreme Gradient Boost had an area under the ROC curve of 87.5%, 80.8% accuracy, over 90% sensitivity, and over 80% specificity, making it the best model for binary classification. In terms of regression analysis, we discovered that the Deep Neural Network had a Pearson coefficient of 0.97 and an R-squared value of 0.94. A family history of autoimmune disease, a personal history of breast cancer, surgically induced menopause, anemia, hypertriglyceridemia, hyperuricemia, fasting hyperglycemia, and elevated alanine aminotransferase levels were all confirmed by statistical indicators used for the regression part of the study as significant risk factors for Hashimoto's thyroiditis.

Conclusions. The suggested machine learning models are effective for diagnosing Hashimoto's thyroiditis when combined with multiple factors. These findings advocate for screening for autoimmune thyroid disease in people with metabolic syndrome, breast cancer patients, and in women with surgically induced menopause.

ABSTRACT. Animated reconstructions are problematic, and the best results arrive from multiple cameras. Most of our video content is monocular, and our experiment takes as input a monocular video and outputs a graph of RGBD images used to create a continuous RBGD video of a target character. Our main contribution relies on creating a database of sprites from the video, treating each sprite as a partial pixel point cloud, and then applying point cloud registration to transition from one frame to another fluidly. For this experiment, we selected video footage of a football match. We used this technique for an individual player - given that extracting continuous sequences of sprites is more challenging and the animations are much more complex. To evaluate, we show that frame-to-frame similarity is better with our proposed technique by computing a minimum spanning tree as transition cost between different characters poses projected back through splatting.

ABSTRACT. In this paper, we show that searching for good Artificial Neural Network (ANN) topologies can be done by Genetic Algorithms (GA) using very small populations, as low as 2 individuals.

We use an Island Model Memetic Algorithm, where evolution optimises the type, number and size of layers, and the weights are adjusted through Backpropagation. We test this algorithm on the MNIST dataset. Although we start on population sizes as large as 125 (5 islands of 25 individuals each), and obtain the best results (96%-97% accuracy) on a population size of 100 (10 islands of 10 individuals each), GAs with population sizes as low as 2 use 680 fitness evaluations to produce ANNs reaching, in the best case, 96%-97% accuracy.

The main drawback is that low population sizes induce result instability: in the average case, the accuracy is between 92% and 95%.

We explain the results, in part, as due to representing layers as a variable-length list - which allows for some flexibility yet restricts the network topology search space - and genetic operators customised for specific layers.

ABSTRACT. Evolutionary Art is a type of generative art which evolves until it achieves the desired form. Our approach to Evolutionary Art is a generator in Minecraft that tries to recreate a cuboid from a hand-drawn approximation of five of its faces.

We try to approximate the target 32 x 32 x 32 cuboid from its provided face projections, and use edge detection to further extract possible "intent" from the shape boundaries in the target images.

Our approach achieved good results on some of the targets we created, while it struggled with one of the targets containing small detail. However, these results are purely subjective since art is, by itself, subjective.

Our aim was not to fully approximate the target; we could achieve this goal with relatively simple deterministic algorithms. We relied on the Genetic Algorithm's global search as a source of "creativity", and sought to obtain varied structures somewhat similar to our pre-drawn target image.

ABSTRACT. For engineers to create durable and effective electrical assemblies, modelling and controlling heat transfer in rotating electrical machines (such as motors) is crucial. In this paper, we compare the performance of three multi-objective evolutionary algorithms, namely NSGA-II, NSGA-III, and MOEA/D in finding the best trade-offs between data collection costs/effort and expected modelling errors when creating low-complexity Linear Regression (LR) models that can accurately estimate key motor component temperatures under various operational scenarios. The algorithms are integrated into a multi-objective thermal modelling strategy that aims to guide the discovery of models that are suitable for microcontroller deployment. Our findings show that while NSGA-II and NSGA-III yield comparably good optimisation outcomes, with a slight, but statistically significant edge for NSGA-II, the results achieved by MOEA/D for this use case are below par.

ABSTRACT. Improving performance of ensemble models can be achieved through the use of model compression. One specific case of model compression is network distillation which aims to transfer the knowledge from a large model to a smaller model. Dataset distillation is an orthogonal task which aims to synthesize a small dataset, encapsulating the implicit knowledge of the entire training dataset, with the goal of obtaining a model, via a ML algorithm, which would perform similarly to a model trained on the original dataset. Existing literature proposes several data distillation techniques involving prototypes generation, usage of gradient based optimizations algorithms, or evolution strategies. The current paper proposes a technique for dataset distillation based on multiobjective genetic algorithms with variable length chromosomes. The images to be distilled are encoded into variable length representations which, in turn, are processed by the Genetic Algorithm over several epochs. Experimental results suggest that the algorithm produces good results in comparison with baselines and indicate potential for further improvements leveraging the versatility of genetic algorithms.

ABSTRACT. This study aims at obtaining a classification of the optimization landscapes for any function, via features of the footprints of simulated annealing (SA) runs on the respective function. Two types of classification are investigated: function-classification (classes labeled by the names of wellknown test functions) and landscape-type-classification (classes labeled by characteristics of the landscape). The idea behind our approach is that an SA-controlled dynamic balance between exploration and exploitation produces during the runs, through the candidate solutions, probability distributions which capture relevant information on the search space landscape; we use three such distributions. In order to achieve comparable SA behaviour across multiple functions, a monotonic decrease of the expected worse-candidate acceptance probability is enforced, with its value reaching 0 exactly at the end of the optimization process. The present study empirically shows the viability of this feature identification technique for various classification tasks on unknown functions.

ABSTRACT. Eye detection represents an essential facial feature to be detected in driver monitoring systems representing the basis for further processing for attention or drowsiness detection. The machine learning approach for infrared image vision problems comes with multiple obstacles like creating a recordings data set with drivers and labeling it to be able to train a model. In very strict areas like automotive or medical imaging machine learning approaches are still a big debate especially because of the black box that is represented by the model, meaning that a wrong detection is impossible to be predicted or explained. That's why the entire focus is shifted to control the data set and the labeling process with very hard manual effort. This paper presents the experimental results of training convolutional neural networks for eye detection in automotive industry using ground truth data obtained from an automatic labeling module. These results are compared with neural networks that were trained using the same data set but with labels created by manual human effort. This experiment shows that similar accuracy of convolutional neural network can be obtained in image vision problems without manual work for image labeling.

ABSTRACT. In this paper, it will be discussed about the development of a toolchain for integrating a convolutional neural network into a car’s cabin, with emphasis on the compiler. The aim of the project is enabling the possibility of introducing the presence of an AI into the cabin, that would mainly supervise the driver for the sake of lowering the rate of any unfortunate event that is happening on the roads while driving, like for example noticing and alerting the driver of sleepiness, dizziness. It may also supervise any other passenger as well. The compiler being used is LLVM, a project aiming to create a modular compiler, dividing itself into 3 main sections: the frontend module, the optimizer and the backend module. The current setup will be using the ONNX-MLIR project as a frontend module because it is intended that an ONNX trained CNN model will be used. As a backend module, an extended RISC-V module will be developed, because the company where the project is being developed, aims to create it’s own hardware accelerator. Thus, the following aspects can be read in this paper: the implementation details about extending the RISC-V backend module for one’s needs, additional information about integrating a virtual hardware simulator into a development environment, namely the Comet RISC-V simulator.

ABSTRACT. LLVM is an open-source collection of compiler and tool-chain technologies designed to be modular and reusable. Its goal is to provide a modern Static Single Assignment compilation strategy capable of supporting the both static and dynamic compilation of arbitrary programming languages. ONNX is an open format built to represent machine learning models. ONNX defines a common set of operators, the building blocks of machine learning and deep learning models, and a common file format to enable artificial intelligence developers to use models with a variety of frameworks, tools, runtimes, and compilers. RISC-V is an open standard instruction set architecture (ISA) based on established reduced instruction set computer (RISC) principles. This paper presents an implementation of a backend instruction for compiling Reshape operator from ONNX machine-learning standard. This instruction has the purpose to compile the operator and generate low-level code specific for RISC-V architecture.

ABSTRACT. Effective traffic management is a vital component in ensuring the smooth and efficient movement of vehicles, reducing congestion, and enhancing overall transportation systems. As urban populations continue to grow, the challenges associated with traffic congestion, environmental impact, and safety concerns become more pronounced. Therefore, the need for innovative and intelligent traffic management solutions is imperative. One key aspect of traffic management is traffic signal optimization. Traditional fixed-time signal systems can be replaced by dynamic adaptive systems that continuously monitor traffic patterns in real-time and adjust signal timings accordingly. By dynamically responding to traffic conditions, these systems optimize traffic flow, reduce delays, and minimize congestion. This paper propose a traffic management system that address most of the traffic congestion problems. The proposed system itself attempts to benefit on already existing V2V and V2I communication and provides an IPC based system with multiple clients, servers and proxies that can achieve data collection and processing on a global scale. Also, the system bases on 8 traffic states that in normal conditions follow one after the another in a cycle, just like with any other junctions at the given moment, but the normal flow can be altered in specific scenarios to improve traffic flow. Unlike other intelligent traffic systems presented in the literature, that try to predict the traffic, a simpler, more effective approach was provided that dynamically adapts based on real-time traffic conditions. The green light time and waiting time are determined based on the number of vehicles that crossed or are waiting at the junction.

ABSTRACT. Due to their natural and inherent way of capturing concurrency, dataflow descriptions of stream programs have seen prevalent usage in fields such as video processing, networks and scientific computing, where they are deployed on many-core, heterogeneous and distributed architectures. Despite robust research on the topic, obstacles still exist in evaluating the performance of stream programs accurately, especially without a complete implementation down to the selected platforms. In this work we introduce and provide a proof of concept for an automated design space exploration flow where causation traces and simulation are used to inform the mapping and scheduling of stream programs running on distributed platforms, using only high-level models of the architecture. The basic idea behind the flow is to profile the designs under well performing mappings and schedules very early in the design process to accurately gauge performance potential.

ABSTRACT. In recent works we have introduced the quantitative programming (or performance evaluation programming) paradigm which provides a framework for the formal verification of concurrent programs. By partitioning the set of program states into bisimulation equivalence classes, this paradigm enables the formal verification of concurrent programs with large state spaces. For formal verification, (as in our previous works) we employ probabilistic model checking techniques. To enable formal verification, the programmer must bound the ranges of variables, and programs are translated into corresponding finite state probabilistic models. In this paper we introduce new quantitative programming primitives that enable the programmer to maintain the compliance between the program and the corresponding probabilistic model, taking into consideration the execution rates of program statements. For formal verification we use Continuous Time Markov Chains and a notion of strong bisimulation specific to stochastic process algebras. We present formal verification experiments that were performed using the PRISM probabilistic model checker.

ABSTRACT. In the link-register model, there is a delay between the time an action is taken and the time an adjacent node is informed of the resulting modification. This delay allows for the study of the asynchronism induced by communications in distributed systems. Read/write atomicity is the most restrictive model in this category, allowing only node-to-node communications. The unfair distributed daemon tops it off by being able to postpone a communication for arbitrarily long (precisely, until the algorithm can take no other move).

This paper proposes a transformer to convert a self-stabilizing algorithm from the state model to the link-register model. In the worst case, one move of the self-stabilizing algorithm in the state model can generate $\Delta$ rounds in the transformed algorithm in the link-register model (where $\Delta$ is the maximum degree of the graph). This transformer is based on another transformer that goes from the state model to a slightly modified version of the link-register model, called the strong-link-register model, in which a node can read in its own registers. This transformation comes with a $O(\Delta)$ factor cost.