Seminars of the Postgraduate Program in Informatics at UnB

Date : June 2, 2023

Time: 2pm

Speaker: Nilson Donizete Guerin Junior   (doctorate degree)

Advisor: Prof Bruno Macchiavello

Title: Neural Image Coding

Summary: Rate control is a desirable and, in some circumstances, necessary feature for image encoding. In image coding, the objective in this case is usually to achieve a desired rate for each encoder input, so that there is minimal impact on rate-distortion performance. This task, therefore, can be quite challenging. In the context of image coding, a new paradigm, based on the use of neural networks, has been adopted as a tool to leverage results in the area of image compression. However, the main approaches in the literature require the training of several neural models for different quality and rate requirements. Even when this characteristic is achieved using a single model, there is no control over the rate operating points obtained. Therefore, a coding tool that provides the ability to obtain specific rates in neural networks can be a determining factor in enabling the adoption of this paradigm in coding standards. With this in mind, the objective of this work is to propose approaches that make neural models obtain specific rate operating points, seeking to minimize quality losses. Results and analyzes presented show the feasibility of this objective.

Time: 2:20 pm

Speaker: Flávia Maria Alves Lopes   (doctorate degree)

Advisor: Profa Alba Melo

Title: An analysis of phylogenetic tree construction methods

Summary: In this lecture, we will present the state-of-the-art in the construction of phylogenetic trees by discussing works that use two methods and we will end with a comparative analysis of the main solutions in the area. The inference of phylogenetic trees can be done using the distance-based method or the character-based method. The distance-based method needs to perform the calculation between pairs of genetic sequences and the distance resulting from their comparison is used to build the phylogenetic tree. The character-based method, which uses inference methods, includes the use of Bayesian inference, maximum parsimony and maximum likelihood .

Time: 2:40 pm

Speaker: Rubens Marques Chaves   (master's degree)

Advisor: Prof. Luis Paulo Faina

Title: Financial Distress Prediction in an Unbalanced Data Stream Environment

Abstract: Corporate bankruptcy predictions are important to companies, investors and authorities. However, most bankruptcy prediction studies have been based on stationary models, therefore they tend to ignore important characteristics of financial distress data like non-stationarity, conceptual drift and data imbalance. To overcome them, this study proposes to identify the most appropriate techniques for dealing with these data characteristics in financial statements provided quarterly by companies to the Securities and Exchange Commission of Brazil (CVM). Experiments were carried out on a sample of data collected from the CVM open data portal, over a period of 10 years (2011 to 2020), with 905 different corporations, 23,834 records with 84 indicators each. In these experiments, a sliding window and a forgetting mechanism were employed to avoid the degradation of the predictive model. The majority of samples have no financial difficulties and 651 companies have financial difficulties. Due to characteristics of the problem, especially the data unbalance, the performance of the models were measured through AUC, G-mean and F1-Score and archived 0.95, 0.68 and 0.58, respectively.

Time: 3pm

Speaker: Lindeberg Pessoa Leite   (doctorate degree)

Advisor: Prof. Teófilo de Campos

Title: Improving linear B-cell epitope prediction by transferring learning from higher to lower taxonomic levels

Abstract: Identification of linear B-cell epitopes (LBCEs) plays a key role in the development of diagnostic tests and vaccines against infectious diseases. However, experimental methods used to determine LBCEs are costly and time-consuming. This has motivated the development of computational methods for the rapid identification of LBCEs based on protein sequence data. To date, multiple machine learning approaches have been developed to address this task. These methods rely on having access to a sufficient amount of epitope data for training the models to predict LBCEs for a specific target organism. However, this data may be limited, especially for less studied organisms. Additionally, these methods face difficulties when dealing with novel pathogens due to the lack of samples in current data bases. In this work, we show how transferring learning from higher to lower taxonomic level can yield substantial performance gains across several quality metrics. By transferring the features learned from organisms that share common ancestors, the method can achieve better performance in predicting linear B-cell epitopes. This leads to increased performance in comparison to state-of-the-art methods for LBCE prediction in terms of F1 and AUC scores. The results suggest that transfer learning across taxonomic levels can significantly enhance the classification of linear B-cell epitopes, which can be particularly useful in predicting epitopes in new and less studied pathogens with limited training data availability.

Time: 3:20 pm

Speaker: Igor da Silva Bonomo (PhD)

Advisor: Prof Eduardo Alchieri

Title: Use of Nilcatenation Graphs in Identification of Transactions for Pruning in Blockchains

Summary: Blockchain technology has consolidated itself with its use in various areas of knowledge, with its greatest use in cryptocurrencies. However, there are still limitations to its wide use in various applications. A relevant limitation is the growing size of blockchains, which causes both storage problems and an increasing need for node synchronization time. In this context, this article studies a solution for pruning in blockchains. This technique consists of removing transactions from the network without harming the consistency of the blockchain. The proposed solution is based on nilcatenation graphs, whose objective is to identify a subgraph that can be removed without compromising the consistency of the network. Experimental results show that the proposed solution can more accurately identify transactions that can be removed (reaching a reduction of up to 20% of transactions), when compared to current techniques (managed to provide up to 5% reduction) that seek to find cycles of transactions in these graphs that can be removed.

Location: Teams- Team PPGI0095 Seminar, Channel 1-2023

PPGI0095 Team

Prof. Célia Ghedini Ralha ( This email address is being protected from spambots. You need JavaScript enabled to view it. )

Coordinator Postgraduate Seminars in Informatics 1-2023