ABSTRACT. The "Evolution of Mona Lisa" problem aims to approximate a target image by overlapping many semi-transparent polygons. The problem has been tackled in the past using multiple Nature-Inspired heuristics, and our main contribution is adding image-centric mutation operators (scaling, rotating and translating polygons). We compare Genetic Algorithms, Hill-Climbing and Simulated Annealing. A candidate solution has variable length (of, at most, 300 decagons) and, due to the variable opacity of polygons, order matters -- resulting, in practice, in a kind of messy GA. We use the same representation and mutation operators for the trajectory methods which, due to the focus on wall-clock time, outperform our GA implementation. We find that these methods retain good image approximation at good run times: 98.9-99.2\% (mean on 30 images), with a time limit of 30 minutes, on images 500-pixels tall.

ABSTRACT. Breast cancer is the most commonly diagnosed type of cancer. It is essential to classify patients as quickly as possible into groups with a high or low risk of cancer, to provide adequate treatment. This paper aims to address the impact of the parameters of convolutional neural networks in the binary classification of mammograms. In this paper, we treat two types of binary classification, namely: classification between normal and abnormal tissues, respectively classification between benign and malignant tumors. In the analysis, we investigate the correlation and impact of batch size and learning rate in increasing the performance of the proposed model. Following the experiments on the MIAS dataset, we concluded that for the treated problems, it is appropriate to choose a learning rate lower than 0.001. For the classification of tissues (normal/abnormal), we obtained the fact that training the model on a batch size of 32 brings the best results, namely an accuracy of 0.67, and for the classification of tumors (benign/malignant), it is more appropriate to use a batch size of 8, for which we obtained an accuracy of 0.63. For the best results configurations, we continued the experiments by investigating the impact of data augmentation. We have increased the number of training data by applying horizontal flip and rotation operations. Following these attempts, we noticed an improvement only for the tissue classification, for which we obtained an accuracy of 0.70.

ABSTRACT. Densest $k$-subgraph problem is a relaxation of the well known maximum clique problem and consists in finding a subgraph with exactly $k$ nodes having a maximum number of edges. An Ant Colony Optimization based approach is proposed to solve this combinatorial optimization problem. Numerical experiments show the effectiveness and potential of the proposed approach.

ABSTRACT. The classification and at the same time the interactive characterization of both bad connections, called alerts or attacks, as well as normal connections, is vital for monitoring network traffic. For this specific task, we developed in this study a neuro-symbolic predictive model based on Logic Tensor Networks. Moreover, we present in detail the advantages and disadvantages of using our hybrid system versus the usage of a standard feed-forward deep neural network classifier. For the comparison to be relevant, the same dataset was used during training and the metrics resulted have been compared. A review of the resulted metrics shows that while both algorithms have similar precision, the hybrid approach gives also the possibility to have interactive explanations and deductive reasoning over data.

ABSTRACT. Machine Learning systems based on deep neural networks are powerful tools, but their wide adoption has shown that both the designers and the users of DNN models must fight the barriers of understanding and controlling what has been learned from data. To make such systems explainable and interpretable, model-specific post-hoc methods have been developed in the literature. This paper presents a family of such methods, Class Activation Mapping, used to explain the image classification process in Convolutional Neural Networks, and achieves a thorough evaluation of these methods. The analysis is done both from a qualitative point of view, through the visualization of the activation heatmap of the image, and from a quantitative point of view, through several metrics that try to objectively quantify the relevant parts of an image that contributed to the classification. Several datasets are used to evaluate the discussed methods and a comparison between the obtained results is presented.

ABSTRACT. We present an approach for automatic fact checking, given a trusted knowledge base and a natural language text. The FACE (FAct Checker with Explanations) system is capable of extracting the knowledge behind the sentences, and decide what is entailed in the trusted sources and what is in conflict with them, providing also explanations and counter speeches in English. The system also specifies the provenance of each of its argument, thus it can be traced back to the source of the information.

Description logic representation of the input is obtained using the FRED machine reader, which is further improved by detecting and handling translation patterns. The obtained ontology is aligned to the knowledge base using the WordNet database in a custom algorithm, then entailment and conflict detection is performed with the Hermit reasoner, through which we obtain the explanations and counter speeches which are verbalized to Attempto Controlled English.

The fact checker is demonstrated on Covid-19 related sentences, however it is domain independent, and can be used with other knowledge bases as well.

ABSTRACT. Automatically mapping wounds from a 2D diagnostic photo onto a 3D morphable model benefits the treatment of acute burn wounds and the documentation of chronic wounds in that it makes the annotation process faster and reduces human error. We propose a pipeline comprising patient shape estimation, wound segmentation, and 2D to 3D location mapping based on deep learning (DL) techniques that facilitates automatic wound transfer from 2D photos onto 3D models.

ABSTRACT. Most of intrusion detection systems nowadays have an analysis based on signatures in place that often fails for newer or modified malware versions. Intrusion detection systems working with cryptanalysis would offer some advantages against obfuscated code or newly derived viruses based on classic exploits. In this paper, we are applying an index of coincidence approach from cryptanalysis with a N-gram pattern-matching technique on recent binaries, to attempt classification of malicious code. Those characteristics are studied with the use of modern Data Mining methods, namely K-means, to discover interesting clusters for classification and properties of malicious behavior. The challenges of gathering, working with, learning from and classifying large amounts of virus data sets with different techniques are explored.

ABSTRACT. Traditionally, enterprises had relied on information systems (IS) as a medium to represent reality, but recently, there is a growing trend among IS scholars admitting that contemporary environment is calling for a change in the view about the IS’ implications in our lives. They acknowledge that our world is increasingly shaped by what is created and performed in variate digital environments. In this view, we ask how conceptual modelling (as a particular manifestation of IS) could be used to shape this new environment while continue being used as a means to represent the domain of discourse. This work at hand presents a proof-of-concept for the new enterprise systems where conceptual modelling plays a key role and the modelling tool acts as a transition layer between the physical and digital realities.

ABSTRACT. The constant increase of published papers and participating authors in conferences has risen the interest in studies related to conference analysis techniques. In the literature, there are numerous different such studies, which offer distinct perspectives regarding the author ranking, impact, collaborations, research interests and many more. In our research, we are introducing a new collaborator group ranking analysis based on four different centrality measures, using published papers data from the SYNASC conference during the 2005-2021 period with which we construct a co-authorship network. The results obtained by using this approach are highlighting the different rankings of the authors and collaborator groups from different perspectives and their overall impact on the conference.