previous day
all days

View: session overviewtalk overview

10:00-10:50 Session 18: Industrial Session
Carbon footprint for training deep learning models on accelerated AI systems

ABSTRACT. The presentation will cover research on the AI model's training impact on the environment (CO2e levels). Besides, we will address various techniques to address a lower CO2e when training deep neural networks.

AI – coming of age

ABSTRACT. The Internet of Things (IOT) has made possible for the past decade several smart-use scenarios that relied on a wireless network of sensors. It has been already deployed successfully in consumer, commercial, industrial, and infrastructure spaces. However, its use has posed significant limitations for privacy and security domains. To solve all this more computing power had to be added, within a power envelope, and with as little network touch as possible. The advent of AI (Artificial Intelligence) has solved these limitation by implementing the "Edge AI" concept, which is actually AI inference on the edge, in the field. This is in a nutshell the AI IOT made possible with special computer architectures dedicated to neural network inference functionality. 

Intel has acquired 3 years ago Movidius to make these new markets possible. As a result, Edge AI is coming out of age and it is showing the first really smart applications. This presentation will list such success stories and use scenarios

11:10-12:50 Session 20: Artificial Intelligence (3) & NCA workshop
An analysis of aggregated coupling's suitability for software defect prediction

ABSTRACT. Software Defect Prediction is an important problem during the development of a software system, because it helps to focus the testing effort on those parts of the system which have a high probability of being defective. It is also well-researched, there being many papers presenting Machine Learning-based prediction models for this problem. But most of them use the same object-oriented structural software metrics as features. In this paper we investigate the impact of aggregated coupling, which combines structural and conceptual coupling, on software defects proneness. In this regard, we present three software metrics suites derived from both structural and conceptual coupling and analyze how their different combinations influence the performance of software defect prediction models. We analyze the relative performance of the models when using features extracted with LSI versus Doc2Vec in conjunction with Cosine versus Euclidean similarity for computing the conceptual coupling. The results suggest that all these features are complementary and their usage improves the performance of the machine learning models.

Hybrid Hyper-parameter Optimization for Collaborative Filtering

ABSTRACT. Collaborative filtering (CF) became a prevalent technique to filter objects a user might like, based on other users' reactions. The neural network based solutions for CF rely on hyper-parameters to control the learning process. This paper documents a solution for hyper-parameter optimization (HPO). We empirically prove that optimizing the hyper-parameters leads to a significant performance gain. Moreover, we show a method to streamline HPO while substantially reducing computation time. Our solution relies on the separation of hyper-parameters into two groups, predetermined and automatically optimizable parameters. By minimizing the later, we can significantly reduce the overall time needed for HPO. After an extensive experimental analysis, the method produced significantly better results than manual HPO in the context of a real-world dataset.

A Sentiment-based Similarity Model for Recommendation Systems
PRESENTER: Sergiu Limboi

ABSTRACT. Recommendation Systems are tools that interpret the users' preferences in an attempt to generate fitting suggestions. Studies in this domain of research tend to conclude that the numerical user ratings are not powerful enough to truly express the users' preferences. The best way to overcome this is by extending the analysis to other elements provided by the user, such as text-based reviews of items. This data is believed to reveal a deeper understanding of the user's sentiment regarding a certain item.

The goal of the proposed paper is to exploit the valuable information offered by the textual reviews, by mixing Sentiment Analysis techniques into the recommendation process.

The contributions of this paper bring two major improvements to the traditional k Nearest Neighbors collaborative filtering algorithm. As a first step, a sentiment rating approach is developed based on calculated sentiment scores for each item. The resulting sentiment ratings replace the numerical ones in the recommendation process. Next, a sentiment based user similarity measure is defined taking into account three factors of similitude: the attractiveness, relevance, and popularity of reviews and users.

Several experimental setups using two different datasets demonstrate that the newly proposed similarity measure outperforms some of the traditional ones and can be successfully used in the recommendation process.

Increasing the Upper Bound for the EvoMan Game Competition

ABSTRACT. This paper describes a comparison between algorithms for evolving agents able to play the game Evoman. Our team took part in the "Evoman: Game-playing Competition for WCCI 2020", and won second place; beyond finding a good agent to satisfy the requirements of the competition - which aim at a good ability to generalise -, we have surpassed the existing non-general, best-known upper-bound. We have managed to exceed this upper bound with a Proximal Policy Optimization algorithm, by discarding the competition requirements to generalise. We also present our other exploratory attempts: Q-learning, Genetic Algorithms, Particle Swarm Optimisation, and their PPO hybridizations. Finally, we map the behaviour of our algorithm in the space of game difficulty, generating plausible extensions to the existing upper-bound.

Experimental Study on Parallelization of Metaheuristic Algorithms
PRESENTER: Csaba Püsök

ABSTRACT. In this work are proposed some parallelization methods of three metaheuristics: Firefly Algorithm, Particle Swarm Optimization and Differential Evolution. A comparation of these parallelized methods is done from precision of the solution and execution time points of views. These metaheuristics are used in our work in order to perform function optimization which is also known as mathematical optimization. Experimental results, conclusions and future work are also presented in this paper.

14:00-15:40 Session 21: Workshop - Big Data Applications (BiD)
Big Data Lab presentation
PRESENTER: Alin Semenescu

ABSTRACT. The multidisciplinary laboratory for Big Data Science seeks to capitalize on the existing interest in several scientific fields for the Big Data field, through a large-scale approach, at the level of the entire West University of Timișoara and to prevent the fragmented distribution of resources and enhance the institution's capacity to capitalize on the results. The latest scientific achievements of the lab members will be presented shortly. 

Optimizations for Deep Learning Models for Object Detection

ABSTRACT. Deep learning models for object detection have gotten larger and larger over the years, spanning from 3.9M trainable parameters for EfficientDet to 209M for the AmoebaNet-based NAS-FPN detector. Different strategies are currently being researched in order to improve the efficiency of deep learning models for object detection, one of which is running the training and inference of the neural network in low precision. Interesting results have been achieved by researchers, starting from the original paradigm of using operators and doing the necessary operations in IEEE single precision (FP32), to achieving similar accuracies of the models using custom minifloat formats (FP8). The results can be pushed even further by using genetic algorithms for hyperparameter tuning, in order to find specific hyperparameter for the FP8 version of the model. In this papers, We will present the results of our experiments utilizing YOLOv3 with hybrid floating-point format (HFP8).

A bibliometric overview of the International Symposium on Symbolic and Numeric Algorithms for Scientific Computing between 2005 and 2018

ABSTRACT. Current research offers a bibliometric overview of the International Symposium on Symbolic and Numeric Algorithms for Scientific Computing, from 2005 to 2018, from different perspectives, in order to highlight the generated impact, the dimensions and strength of international collaborations, as well as a statistical study of conference papers, typical structure of collaboration groups, evolution of research trends, and others. Associated findings are presented either as raw data, or processed via VOSViewer.

Detecting Individuals High in Neuroticism based on the Color Features of the Facebook Profile Picture

ABSTRACT. Previous research has mostly focused on the link between the linguistic and behavioral footprints found on social media on the one hand and personality on the other. Despite the high amount of image-based contents posted and shared online and the valuable implicit information they might conceal about users’ preferences and tendencies, the study of visual traces is in its infancy. The goal of the current paper is to test whether the color characteristics of the Facebook profile picture could mirror the level of neuroticism on a sample of 508 Romanian users. For this purpose, we assessed the classification performance of four machine learning algorithms having as input three sets of visual features: (1) the colorfulness, which indicates how colorful is an image; (2) the proportion of cold colors, along with the mean and standard deviation for saturation and value; (3) the emotional load, defined as pleasure, arousal, and dominance. None of the models showed good accuracy. However, this paper contributes to the literature by being part of a line of research that requires development not only in Romania but also worldwide.

A dynamic contextual citation graph of academic papers

ABSTRACT. Nowadays, citations have become an important source of information. We face the need to structure our knowledge and find the right context of our research. This paper presents a software system that fulfills this need by building a graph in which we link academic papers based on citations between one another. Its dynamic ability comes from the fact that once a new paper is submitted, it can automatically find dependencies to other papers and vice versa. The connections between the nodes in this graph are based on relevant information about it, such as, complete author names, title and citations that are used within. A paragraph in which a citation is used is also extracted from the paper in order to have more information about citation context. The result of this processing is converted into our own model which is then stored in a graph database. Using the query language, various information such as ”what is the most cited paper and in what context do these citations appear” can be retrieved and viewed.

16:00-16:50 Session 22: Invited talk
Operational Semantics and Program Verification using Many-sorted Hybrid Modal Logic

ABSTRACT. In this talk we present the many-sorted hybrid modal logic, a logical system that is powerful enough to represent both the programs and their semantics in a uniform way. This logic is built by performing hybridization on top of a general many-sorted polyadic modal logic. Given a propositional modal logic, its hybrid companion is defined by adding nominals, atomic symbols that name the states of a model, as well as special operators and binders. The many-sorted hybrid modal logic is a sound and complete system with the property that, once a language is specified, one can define its operational semantics and perform Hoare-style verification. The SMC-inspired operational semantics is defined by a set of axioms derived from those of Propositional Dynamic Logic, and general Hoare-like assertions can be proved in this setting. We present our approach from general to particular, as well as challenges and future work.

17:10-18:30 Session 23: Logic and Programming (2)
Fischer-Ladner Closure for Many-sorted Modal Logic with Application for Operational Semantics

ABSTRACT. In prior work we have developed a many-sorted polyadic modal logic. Progressively adding different operations and binders, we expressed operational semantics, thus enabling us to certify execution. In this paper we prove standard completeness for a many-sorted hybrid modal logic with satisfaction operators and PDL-inspired axioms. In order to do this, we define the Fischer-Ladner closure for our particular system and we prove a variant of the small model property.

On One Approach to Goal-driven Proof Search in Classical First-order Logic

ABSTRACT. An approach to a computer-oriented proof search in classical first-order logic based on goal-driven sequent calculi that do not require skolemization and one of which is a quantifier-rules-free one is developed. Results about the soundness and completeness of proposed sequent calculi are obtained through establishing their coextensivity with a usual Gentzen-type sequent calculus.

Functional Programming GUIs state of the art

ABSTRACT. In recent years versions of a functional approach to the problem of programming GUIs -- namely Functional Reactive Programming -- have known a great deal of success being implemented in imperative languages mainly due to providing a good structure to Single Page Applications written in JavaScript. This is an indication that there is a great interest in the simplicity provided by the functional framework to the overall process of structuring an application.

As popular as the approach is it still does not fulfill the age-old promise of perfect decoupling of components in programming nor the one of minimizing the cognitive burden on the programmer when trying to achieve this decoupling. Moreover, complexity appears to be the single most frequent cause of software errors, leading to less stable software. Due to concurrency this issue is especially prevalent in GUI programming.