Brazil

Stochastic Shortest Path(SSP) is the most popular framework to model sequential decision-making problems under stochasticity. However, decisions for real problems should consider risk sensitivity to provide robust decisions taking into account bad scenarios. SSPs that deal with risk are called Risk-Sensitive SSPs (RSSSPs), and an interesting framework from a theoretical perspective con- siders Expected Utility Theory under an exponential utility function. However, from a practical perspective, exponential utility function causes overflow or underflow in computer implementation even in small state spaces. In this paper, we make use of LogSumExp technique to solve RSSSPs under exponential utility in practice within Value Iteration, Policy Iteration, and Linear Programming algorithms. Experiments were performed on a toy problem to show scalability of the proposed algorithms.

BRACIS 2020

Risk Sensitive Stochastic Shortest Path and LogSumExp: from theory to practice

exponential utility

rsmdp

rsssp

The 9th Brazilian Conference on Intelligent Systems (BRACIS), the 16h Encontro Nacional de Inteligência Artificial e Computacional (ENIAC), and the 12th Concurso de Teses e Dissertações em Inteligência Artificial e Computacional (CTDIAC) was held from 20 to 23 October, 2020.

BRACIS originated from the combination of the two most important scientific events in Brazil in Artificial Intelligence (AI) and Computational Intelligence (CI): the Brazilian Symposium on Artificial Intelligence - SBIA (22 editions), and the Brazilian Symposium on Neural Networks - SBRN (13 editions). ENIAC is a national forum for researchers, practitioners, educators and students to present and discuss innovations, trends, experiences, developments and work in progress in the fields of Artificial and Computational Intelligence. CTDIAC is a satellite event whose goals are to disseminate and award the best dissertations and theses (MSc and PhDs concluded in Brazilian universities from June 2018 to May 2020) in Artificial and Computational Intelligence areas.

In this year, the Symposium on Knowledge Discovery, Mining and Learning (KDMiLe) will be held in conjunction with BRACIS 2020. This sysposium aims at integrating researchers, practitioners, developers, students and users to present their research results, to discuss ideas, and to exchange techniques, tools, and practical experiences – related to the Data Mining and Machine Learning areas.

technical paper

Com o crescimento da computação quântica como um paradigma de processamento de informação, modelos quânticos de redes neurais têm sido propostos para unir os benefícios da computação quântica com os benefícios das Redes Neurais Artificiais (RNA). Existe uma dificuldade de implementar a não-linearidade dos neurônios em computadores quânticos, uma vez que tradicionalmente a computação quântica possui apenas operadores unitários. Há algumas propostas na literatura de modelos de neurônios que simulam essa não-linearidade, mas elas aparecem simulando alguma função não-linear específica, como a função de limiar ou arco-tangente. Há ainda as RNAs que possuem comportamento de memória associativa, fazendo a recuperação de informação a partir de uma entrada igual ou parecida com seus padrões armazenados.
Nesse trabalho, três novos modelos inteligentes quânticos são descritos. O primeiro deles é um modelo de neurônio que executa qualquer função não-linear discreta. O modelo proposto permite que o neurônio possa simular o comportamento dos neurônios clássicos, assim como utilizar dos recursos intrínsecos da computação quântica como superposição e emaranhamento. Esse modelo também permite incorporar informações espaço-temporais. O segundo modelo proposto é uma memória associativa com uso de operadores não-lineares e o terceiro modelo é uma memória associativa com operadores lineares, utilizando os algoritmos quânticos de Grover e transformada inversa de Fourier. Esses modelos de memória permitem reconhecer informações próximas ou iguais ao conteúdo que está na memória e possui custo linear de operação.

Avanços em Redes Neurais Quânticas

Action abstractions restrict the number of legal actions available for real-time planning in multi-unit zero-sum extensive-form games, thus allowing algorithms to focus their search on a set of promising actions. Even though unabstracted game trees can lead to optimal policies, due to real-time constraints and the tree size, they are not a practical choice. In this context, we introduce an action abstraction scheme we call asymmetric abstraction. Similarly to unabstracted spaces,asymmetrically-abstracted spaces can have theoretical advantages over regularly abstracted spaces while still allowing search algorithms to derive effective strategies in practice, even in large-scale games. Further, asymmetric abstractions allow search algorithms to “pay more attention” to some aspects of the game by unevenly dividing the algorithm’s search effort amongst different aspects of the game. We also introduce four algorithms that search in asymmetrically-abstracted game trees to evaluate the effectiveness of our abstraction schemes. An extensive set of experiments in a real-time strategy game developed for research purposes shows that search algorithms using asymmetric abstractions are able to outperform all other search algorithms tested.

Asymmetric Action Abstractions for Real-Time Planning in Extensive-Form Games

Statistical machine learning algorithms usually assume that there is considerably-size data to train the models. However, they would fail in addressing domains where data is difficult or expensive to obtain. Transfer learning has emerged to address this problem of learning from scarce data by relying on a model learned in a source domain where data is easy to obtain to be a starting point for the target domain. On the other hand, real-world data contains objects and their relations, usually gathered from noisy environment. Finding patterns through such uncertain relational data has been the focus of the Statistical Relational Learning (SRL) area. Thus, to address domains with scarce, relational, and uncertain data, in this paper, we propose TreeBoostler, an algorithm that transfers the SRL state-of-the-art Boosted Relational Dependency Networks learned in a source domain to the target domain. TreeBoostler first finds a mapping between pairs of predicates to accommodate the additive trees into the target vocabulary. After, it employs two theory revision operators devised to handle incorrect relational regression trees aiming at improving the performance of the mapped trees. In the experiments presented in this paper, TreeBoostler has successfully transferred knowledge among several distinct domains. Moreover, it performs comparably or better than learning from scratch methods in terms of accuracy and outperforms a transfer learning approach in terms of accuracy and runtime.

Transfer Learning by Mapping and Revising Boosted Relational Dependency Networks

The performance of a meta-heuristic on a search problem relies on the features of the problem instance, meaning that it does not exist a single method that always achieves the best performance. Therefore, approaches that automatically select an algorithm can improve the robustness of an application. Among the possible alternatives, the selection can be done with meta-learning, which aims at mapping the characteristics (meta-features) of problem instances to the performance of a set of algorithms. Some works have proposed the use of Fitness Landscape Analysis (FLA) to characterize the instances, where the structure of a problem is analysed by extracting measures of the search space, normally with a sample of solutions. This dissertation is composed by a series of incremental studies on meta-learning algorithm selection for the Quadratic Assignment Problem. We experimented with a variety of meta-features, giving emphasis on the FLA based metrics, and with a set of classic and novel meta-heuristics. The reported results show that this meta-learning approach is suitable for algorithm selection, yielding good and balanced classification performance with a low addition in computational effort.

Seleção Automática de Algoritmo para o Problema Quadrático de Alocação Usando Meta-aprendizagem e Métricas de Superfície de Aptidão

A classificação de textos tem sido estudada extensivamente nas últimas décadas e grande parte dos trabalhos relacionados ao tema são direcionados a classificação de rótulo único e ao aprendizado offline. Neste tipo de aprendizado, os documentos de texto são associados a apenas um rótulo e devem estar disponíveis com antecedência para o treinamento. Problemas reais de classificação de textos, no entanto, frequentemente envolvem instâncias multirrotuladas, que se tornam disponíveis continuamente e com padrões que mudam ao longo do tempo. Para manipular esses problemas, os classificadores idealmente deveriam ser capazes de predizer múltiplos rótulos para cada documento de texto e de atualizar o seu modelo preditivo de forma eficiente, para ser escalável mesmo com recursos de memória e tempo limitados, e ser rapidamente adaptável às mudanças nos padrões dos dados. Por isso, o aprendizado online e a classificação multirrótulo tem atraído grande interesse de pesquisa, uma vez que existem poucos métodos capazes de abordar os dois problemas simultaneamente e, assim, sempre é necessário recorrer às técnicas de transformação de problemas ou retreinar todo o modelo de predição quando novos documentos de texto se tornam disponíveis. Neste trabalho, é apresentado um método de classificação de textos baseado no princípio da descrição mais simples, que pode ser empregado em problemas de classificação multirrótulo sem a necessidade de transformá-los em problemas de rótulo único. Ele também apresenta a vantagem de considerar a existência de dependência entre os rótulos e de suportar o treinamento incremental naturalmente. O desempenho desse método foi avaliado empregando-o na tarefa de classificação em 15 aplicações de diferentes domínios e o resultado obtido foi comparado com os resultados de outros classificadores de referência na literatura, considerando cenários de aprendizado offline e online. Os resultados obtidos pelo método proposto são muito competitivos com os resultados obtidos pelos métodos estado-da-arte avaliados, uma vez que o ML-MDLText esteve
entre as duas melhores performances em termos acurácia de subconjunto e macro F-medida nos experimentos com aprendizado online.

ML-MDLText: um método de classificação de textos multirrótulo de aprendizado incremental

Communication has become increasingly dynamic with the popularization of social networks and applications that allow people to express themselves and communicate instantly. In this scenario, distributed representation models have their quality impacted by new words that appear frequently or that are derived from spelling errors. These words that are unknown by the models, known as out-of-vocabulary (OOV) words, need to be properly handled to not degrade the quality of the natural language processing (NLP) applications, which depend on the appropriate vector representation of the texts. To better understand this problem and finding the best techniques to handle OOV words, in this study, we present a comprehensive performance evaluation of deep learning models for representing OOV words. We performed an intrinsic evaluation using a benchmark dataset and an extrinsic evaluation using different NLP tasks: text categorization, named entity recognition, and part-of-speech tagging. Although the results indicated that the best technique for handling OOV words is different for each task, Comick, a deep learning method that infers the embedding based on the context and the morphological structure of the OOV word, obtained promising results.

Deep learning models for representing out-of-vocabulary words

In recent years, the advances in technology have produced datasets of increasing size, not only regarding the number of samples but also the number of features. Unfortunately, creating a sufﬁciently large amount of adequately labeled data with enough examples for each class is not easy. Labeling is a challenging, expensive, and time-consuming task. It is usually done manually, which may contribute to the insertion of noise and errors in the data. Hence, it is of great importance to put forward intelligent models that can beneﬁt from the distinct information that both labeled and unlabeled data can provide, since, for many applications, there is a plentiful amount of unlabeled data, but insufﬁcient labeled ones. Semi-Supervised Learning (SSL) is employed to achieve this. It is halfway between supervised and unsupervised learning. In this sense, we highlight two very inﬂuential models: Self-Organizing Maps (SOM) and Learning Vector Quantization (LVQ). SOM is a biologically inspired neural model that uses unsupervised and incremental learning to produce prototypes of the input data, whereas the LVQ can be seen as its supervised counterpart. The unsupervised characteristic of SOM makes it unfeasible to execute SSL. In that way, the current work proposes new models that incorporate standard concepts from LVQ to the SOM algorithm to build semi-supervised approaches. Such proposals can dynamically switch between the two types of learning at training time, according to the availability of labels and automatically adjust themselves to the local variance observed in each cluster. The experimental results show that the proposed models can surpass the performance of other traditional methods not only in terms of classiﬁcation but also regarding clustering quality. It also enhances the range of applications of SOM and LVQ-based models by combining them with Deep Learning in a synergic way to allow dealing with complex data structures, such as images and sound. Moreover, we explore forms of learning good representations of the input data, and manners to estimate the unsupervised error when no labels are provided. Our approaches demonstrated to be good at producing a meaningful topology and clustering prototypes that appropriately represent the data.

Semi-Supervised Self-Organizing Maps with Time-Varying Structures for Clustering and Classification

This work studies quantitative measures for ranking judicial decisions by the Brazilian Supreme Court (STF).
The measures are based on a network built over decisions whose cases were finalized in the Brazilian Supreme Court between 01/2001 and 12/2019, obtained by crawling publicly available STF records.
Three ranking measures are proposed; two are adaptations of the PageRank algorithm, and one adapts Kleinberg's Algorithm.
All are compared with respect to agreement on top 100 rankings; we also analyze each measure robustness based on self-agreement under perturbation.

We conclude that all algorithms show the network of citations is highly robust under perturbation.
Both versions of PageRank, even if producing different rankings, achieved robustness results which are indistinguishable via statistical tests; Kleinberg's algorithm achieves more promising results to rank leading cases at the top, but it does need more research to achieve this goal.


Robust Ranking of Brazilian Supreme Court Decisions

Recent research on Convolutional Neural Networks focuses on how to create models with a reduced number of parameters and a smaller storage size while keeping the model's ability to perform its task, allowing the use of the best CNN for automating tasks in limited devices, with reduced processing power, memory, or energy consumption constraints. There are many different approaches in the literature: removing parameters, reduction of the floating-point precision, creating smaller models that mimic larger models, neural architecture search (NAS), etc. With all those possibilities, it is challenging to say which approach provides a better trade-off between model reduction and performance, due to the difference between the approaches, their respective models, the benchmark datasets, or variations in training details. Therefore, this article contributes to the literature by comparing three state-of-the-art model compression approaches to reduce a well-known convolutional approach for object detection, namely YOLOv3. Our experimental analysis shows that it is possible to create a reduced version of YOLOv3 with 90% fewer parameters and still outperform the original model by pruning parameters. We also create models that require only 0.43% of the original model's inference effort.

An Experimental Analysis of Model Compression Techniques for Object Detection

The world around us is composed of images that often need
to be translated into words. This translation can take place in parts,
converting regions of the image into textual descriptions what is also
known as dense captioning. By doing so, the information present in this
region is converted into words expressing the way objects relate to each
other. Computational models have been proposed to perform this task in
a similar way to human beings, mainly using deep neural networks. As the
same region of the image can be described in several different forms, this
study proposes to use the Abstract Meaning Representation (AMR) in
the generation of descriptions for a given region. We hypothesize that by
using AMR it would be possible to extract the meaning of the text and,
as a consequence, improve the quality of the sentences produced by the
models. AMR was investigated as a semantic representation formalism
evolving the so far proposed models that are based only on purely natural
language. The results show that the models trained with sentences in the
form of AMR led to better descriptions and the performance achieved
was superior in almost all evaluations.

Downloads

Next from BRACIS 2020

Avanços em Redes Neurais Quânticas

Similar lecture

Risk Sensitive Policies for the Control of the Spread of Infectious Diseases

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from BRACIS 2020

Avanços em Redes Neurais Quânticas

Similar lecture

Risk Sensitive Policies for the Control of the Spread of Infectious Diseases

Downloads