Brazil

Forecasting photovoltaic (PV) power generation, as in many other time series scenarios, is a challenging task. Most current solutions for time series forecasting are grounded on Machine Learning (ML) algorithms, which usually outperform statistical-based methods. However, solutions based on ML and, more recently, Deep Learning (DL) have been found vulnerable to adversarial attacks throughout their execution. With this in mind, in this work we explore four time series analysis techniques, namely Naive, a baseline technique for time series, Autoregressive Integrated Moving Average (ARIMA), from the statistical field, and Long Short-term Memory (LSTM) and Temporal Convolutional Network (TCN), from the DL family. These techniques are used to forecast the power generation of a PV power plant 15 minutes and 24 hours ahead, having as input only power generation historical data. Two main aspects were analyzed: i) how training size influenced the performance of the forecasting models and ii) how univariate time series data could be modified by an adversarial attack to decrease models&#39; performance through cross-technique transferability. For i), the mentioned methods were used and evaluated with monthly updates. For ii), Fast Gradient Sign Method (FGSM), along with a logistic regression substitute model and past data, were used to perform attacks against DL models at test time. LSTM and TCN decreased the error as the training sample size increased and outperformed Naive and ARIMA models. Adversarial samples were able to reduce the performance of LSTM and TCN, particularly for short-term forecasts.

BRACIS 2020

Photovoltaic Generation Forecast: Model Training and Adversarial Attack Aspects

smart grid

solar photovoltaic generation

time series forecast

deep learning

adversarial attack

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.

Forecasting photovoltaic (PV) power generation, as in many other time series scenarios, is a challenging task. Most current solutions for time series forecasting are grounded on Machine Learning (ML) algorithms, which usually outperform statistical-based methods. However, solutions based on ML and, more recently, Deep Learning (DL) have been found vulnerable to adversarial attacks throughout their execution. With this in mind, in this work we explore four time series analysis techniques, namely Naive, a baseline technique for time series, Autoregressive Integrated Moving Average (ARIMA), from the statistical field, and Long Short-term Memory (LSTM) and Temporal Convolutional Network (TCN), from the DL family. These techniques are used to forecast the power generation of a PV power plant 15 minutes and 24 hours ahead, having as input only power generation historical data. Two main aspects were analyzed: i) how training size influenced the performance of the forecasting models and ii) how univariate time series data could be modified by an adversarial attack to decrease models' performance through cross-technique transferability. For i), the mentioned methods were used and evaluated with monthly updates. For ii), Fast Gradient Sign Method (FGSM), along with a logistic regression substitute model and past data, were used to perform attacks against DL models at test time. LSTM and TCN decreased the error as the training sample size increased and outperformed Naive and ARIMA models. Adversarial samples were able to reduce the performance of LSTM and TCN, particularly for short-term forecasts.

technical paper

Word segmentation can contribute to improve the results of natural language processing tasks on several problem domains, including social media sentiment analysis, source code summarization and neural machine translation. Taking the English language as a case study, we fine-tune a Transformer architecture which has been trained through the Pre-trained Distillation (PD) algorithm, while comparing it to previous experiments with recurrent neural networks. We organize datasets and resources from multiple application domains under a unified format, and demonstrate that our proposed architecture has competitive performance and superior cross-domain generalization in comparison with previous approaches for word segmentation in Western languages.

Domain Adaptation of Transformers for English Word Segmentation

Restoring corrupted images represent a challenge to computer vision and image processing based on machine learning. In hazy, underwater and medical images the lack of paired images lead the authors to synthesize datasets and the Generative Adversarial Networks (GANs) are widely used. However, computational cost and training instability are current concerns. We present an unsupervised learning algorithm that does not requires paired dataset to train encoder-decoder-like neu-
ral network for image restoration. An encoder-decoder learn to represent
its input data in a latent representation and reconstruct then in the output. During the training stage, our algorithm applies the encoder-decoder output image to a degradation block that reinforces its degradation. The degraded and input images are matched using a loss function. After the training process, we obtain a restored image from the decoder. We used ill-exposed images to evaluate and validate our algorithm.

Unsupervised Learning Method for Encoder-Decoder-Based Image Restoration

Word embeddings is a text representation technique capable of capturing syntactic and semantic linguistic patterns and of representing each word as an n-dimensional dense vector. In the domain of legal texts, there are trained word embeddings in languages like English, Polish, and Chinese. However, to the best of our knowledge, there are no embeddings based on Portuguese (Brazilian and European) legal texts. Given that, our research question is: does the specificity and size of the text corpus used for a word embedding training contribute to a more successful classification? To answer the question, we train word embeddings models in the legal domain with different levels of specificity and size. Then we evaluate their impact on text classification. To deal with the different levels of specificity, we collect text documents from different courts of the Brazilian Judiciary, in hierarchical order. We used these text corpora to train a word embeddings model (GloVe) and then had then evaluated while classifying processes with a deep learning model (CNN). In a context perspective, the results show that in word embeddings trained on smaller corpora sizes, text specificity has a higher impact than for large sizes. Also, in a corpus size perspective, the results demonstrate that the greater the corpus size in embeddings training, the better are the results. However, this impact decreases as the corpus size increases until a point where more words in the corpus have little impact on the results.

Impact of Text Specificity and Size on Word Embeddings Performance: an Empirical Evaluation in Brazilian Legal Domain

An important challenge in the research field of Biometrics isreal-time identification, at a distance, in uncontrolled environments, us-ing low-resolution cameras. In such circumstances, soft biometrics can bethe only option. In this work, we propose two novel descriptor methodsfor biometric identification based on ensemble of anthropometric mea-surements and joints heat-map of the person skeleton, captured fromvideo frames through state-of-the-art 2D poses estimation methods. Theproposed methods were assessed on a popular benchmark dataset, CA-SIA Gait Dataset B, and obtained good results (85% and 89% of rank-1identification rates, respectively) with PifPaf 2D pose estimation method

People Identification Based on Soft BiometricsFeatures Obtained from 2D Poses

Colorectal cancer causes the deaths of thousands of people worldwide according to the World Health Organization. Automatic tissue recognition of histopathological images is essential for early disease diagnosis. Most research consists of employing texture descriptors to capture features that identify tumor samples. However, accurate multi-class classification is a challenge due to the complexity of colorectal tissue images. Recently, researchers have shown that the analysis of texture structural patterns degraded by image filtering provides valuable features for pre-diagnosis in several medical applications. Here we propose an approach to automatically classify eight types of colorectal tissues using Structural Co-occurrence Matrix. We carried on experiments on 5000 tissue patches from a public dataset to evaluate our algorithm, considering two scenarios: structural differences as a single descriptor, and combined with other characteristics. We found that our strategy improves the stateof-the-art, achieving, accuracy: 91.30%, precision: 91.41%, sensitivity: 91.31%, specificity: 98.76% e F1-score: 91.31%.

Texture analysis based on structural co-occurrence matrix improves the colorectal tissue characterization

Early diagnosis of genetic syndromes has a vital importance in the prevention of any potential related health problems. Down syndrome is the most common genetic syndrome. Patients with down syndrome have a high probability of developmental disorders, like Congenital Heart Disease, which is best treated when discovered in the early stages. These patients also have particular facial characteristics that are identified by geneticists in a physical exam. However, there is subjectivity in the professional analysis, which can lead to a late diagnosis, aggravating the patient's health condition. This paper proposes a software framework for the automatic detection of Down syndrome using facial features extracted from digital images, which could be used as a tool to help in the early detection of genetic syndromes. For training the machine learning model, we create a dataset gathering 170 pictures of children available on the internet. 50% of the pictures were of children with Down syndrome and the other 50% of healthy children. Then, we automatically identify faces and describe the images with facial landmarks. Next, we use two approaches for feature extraction. The first is a traditional computer vision approach using selected distances and angles and textures between the landmarks. The other, a deep learning approach using a Convolutional Neural Network to extract the features automatically. Then, the feature vector is fed to a Support Vector Machine with a linear kernel on both feature extraction approaches. We validate the results measuring the accuracy, sensitivity, and specificity of both feature extraction approaches using 10-fold cross-validation. The deep learning method resulted in an accuracy of 0.94, while the traditional approach achieved 0.84 of accuracy in our dataset. The results shows that the deep learning approach has a higher classification accuracy for this task, even with a small dataset.

A computational tool for automated detection of genetic syndrome using facial images

With the abundance of electronic lawsuits already implemented throughout Brazil, courts have a valuable source of information in text format that con-stitute attractive bases for the application of Artificial Intelligence (AI) and machine learning (ML). In this research, supervised learning approaches were explored for the automatic classification of types of documents in elec-tronic court proceedings of the Court of Justice of Minas Gerais (TJMG). The methodology is composed of cross-validation within the specific corpus of the legal domain, comparing traditional classifiers and more recent meth-ods based on neural networks and deep learning models, using Glove word vectors generated for the Portuguese Language and Convolutional Neural Network (CNN). This work achieved high precision in the results and if im-plemented in the courts it can provide significant savings in financial and human resources, allowing lawsuits classification activities, currently done manually by employees, to be performed in seconds by the machine. The re-sult of this experiment shows that the hit rates for the CNN and SVM classi-fiers exceed 93% and is considered a high result. Based on the assumption that Glove brings extra semantic resources that can help in classifying texts from court proceedings, this work demonstrates Glove's effectiveness by showing that a CNN with Glove surpasses SVM.

The use of machine learning in the classification of electronic lawsuits: an application in the Court of Justice of Minas Gerais

ICD coding from electronic clinical records is a manual, timeconsuming and expensive process. Code assignment is, however, an important task for billing purposes and database organization. While many works have studied the problem of automated ICD coding from free text using machine learning techniques, most use records in the English language, especially from the MIMIC-III public dataset. This work presents results for a dataset with Brazilian Portuguese clinical notes. We develop and optimize a Logistic Regression model, a Convolutional Neural Network (CNN), a Gated Recurrent Unit Neural Network and a CNN with Attention (CNN-Att) for prediction of diagnosis ICD codes. We also report our results for the MIMIC-III dataset, which outperform previous work among models of the same families, as well as the state of the art. Compared to MIMIC-III, the Brazilian Portuguese dataset contains far fewer words per document, when only discharge summaries are used. We experiment concatenating additional documents available in this dataset, achieving a great boost in performance. The CNN-Att model achieves
the best results on both datasets, with micro-averaged F1 score of 0.537 on MIMIC-III and 0.485 on our dataset with additional documents.


Predicting Multiple ICD-10 Codes from Brazilian Portuguese Clinical Notes

In this presentation, we will talk about the use of a semi-supervised technique called Randon Walk in Feature Sample Networks and the application of this technique in sentiment analysis of Portuguese tweets. We also present a change in the classification mechanism and the results found.

Semi-Supervised Sentiment Analysis of Portuguese Tweets with Random Walk in Feature Sample Networks

Automatic License Plate Recognition (ALPR) systems are used in many real-world applications, such as road traffic monitoring and traffic law enforcement, and the use of deep learning can result in efficient methods. In this work, we present an ALPR system efficient for edge computing, using a combination of MobileNet-SSD for vehicle detection, Tiny YOLOv3 for license plate detection and OCR-net for character recognition. This method was evaluated in two datasets on a NVIDIA Jetson TX2 system, obtaining 96.87% of accuracy and 8 FPS of framerate in a public real-world scenario dataset and achieving 90.56% of accuracy and 11 FPS of framerate in a private dataset of traffic monitoring images, considering the recognition of at least six characters. It is faster than related works with similar deep learning approaches, that achieved at most 2 FPS, and slightly inferior in accuracy, with less than 10% of difference in the worst scenario. This shows the proposed method is well balanced between accuracy and speed, thus, suitable for embedded devices.

Downloads

Next from BRACIS 2020

Domain Adaptation of Transformers for English Word Segmentation

Similar lecture

Comparative Study of Photovoltaic Power Forecasting Methods

.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

Domain Adaptation of Transformers for English Word Segmentation

Similar lecture

Comparative Study of Photovoltaic Power Forecasting Methods

Downloads