Brazil

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&#39;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&#39;s inference effort.

BRACIS 2020

An Experimental Analysis of Model Compression Techniques for Object Detection

yolov3

object detection

neural architecture search

knowledge distillation

pruning

ICPR 2020 is the premier international conference on Pattern Recognition, gathering leading researchers from around the world to present their latest achievements in Pattern Analysis and Recognition, applied Artificial Intelligence and Machine Learning. The topic of the year is " Putting Artificial intelligence to work on Patterns". 
For the first time, ICPR has employed a two-round review process similar to journal submissions, with new papers allowed to be submitted in either the first or second round and papers submitted in the first round and not accepted allowed to be revised and re-submitted for second round review. 

In this edition of the **ICPR 2020** Conference, there are 1407 Oral/Poster presentations organized in five Tracks: 
- Artificial Intelligence, Machine Learning for Pattern Analysis; 
- Biometrics, Human Analysis and Behavior Understanding; Computer Vision, Robotics and Intelligent Systems; 
- Document and Media Analysis; and 
- Image and Signal Processing. 

The **ICPR 2020** Main Conference program also includes 7 Keynotes from prominent world-renowned scientists, 8 Tutorials, 28 Demos, 8 Challenges, 2 Panels, as well as an Industry/Sponsor Exhibition. In addition, 41 Workshops on a range of important current and emerging research topics complete the ICPR 2020 program. 


ICPR 2020

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.

technical paper

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.

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.

Dense Captioning using Abstract Meaning Representation

How to choose a route that takes you from A to B? This is an issue that is turning more and more important in modern societies. One way to address this agenda is through the use of communication between the infrastructure (network), and the demand (vehicles). In this paper, we use car-to-infrastructure (C2I) communication to investigate whether the road users (agents) can accelerate their learning process regarding route choice problem, via reinforcement learning (RL). We employ a microscopic simulator in order to compare our method with two others: RL without communication and an iterative method. Experimental results show that our method outperforms both methods in terms of effectiveness and efficiency.

Accelerating Learning of Route Choices With C2I: A Preliminary Investigation

This work addresses a critical problem in the use of the Face Recognition by a police department state of Brazil. Great advances have been shown in the last years and the state of art works are based on large labeled datasets and supervised training. But with this approach, the lack of representative data distribution is an issue, known as data bias, mainly according to some aspects that makes FR harders: gender and race. Recent works have suggested that these two aspects may cause a significant drop in accuracy. This work is concerned over the data bias problem for Brazilian faces. Using pre-trained models trained with public datasets, we demonstrate that even with small size dataset it is possible to improve the accuracy of Brazilian faces with simple yet effective implementation tricks in fine-tuning. Two important conclusions was obtained from this study: First, there is a strong suggestion the there is data bias concerning ethnicity when evaluating models on Brazilian faces trained with public datasets, and second, the fine-tuning task using a brazillian non-public dataset resulted in significant mitigation of this problem. 


Improving Face Recognition Accuracy for Brazilian Faces in a Criminal Investigation Department

The internet connection is present in people's lives all the time, through smartphones, tablets, computers, among others. The use of social networks is increasingly common around the world. Many companies and people use them to spread products and services and publish opinions, facts that have turned the social networks into powerful sources of information on various topics. Identifying these feelings is a great strategy for many types of decision making. Thus, the purpose of this paper is to collect messages from a specific social network, in this case Twitter, referring to the 2018 Brazilian presidential elections and classify them as: positive, negative and neutral, in order to discover a possible relationship between opinions of social network users and the final outcome of the elections. For this, a corpus was built, preprocessed and evaluated by two different machine learning approaches: Naive Bayes and SVM (Support Vector Machine). The results showed that this social network is a good source of information to perform sentiment analysis and that the number of tweets classified as positive have a strong relationship with the researchers and the final result of the 2018 elections.

A Sentiment Analysis of Brazilian Elections Tweets

Dependency Parsers (DP) are parsers that analyze dependencies between words in a sentence. Currently, dependency parser evaluation is a problem whose solutions are not well defined in the scientific community. Although the DP intrinsic metrics are the foremost choice of evaluation, extrinsic evaluation enables a different evaluation based on a downstream. Different results of DP can vary according to the domain task. Thus, this work applies an Open Information Extraction (OIE) method in Portuguese to provide an extrinsic evaluation of a set of CONLL Dependency Parsers. Our results demonstrate that there is a difference in the evaluation of Dependency Parsers considering a particular task. 


CoNLL Dependency Parser: Extrinsic Evaluation through the Open Information Extraction task

The lackadaisical quantum walk is a graph search algorithm for 2D grids whose vertices have a self-loop of weight l. Since the technique depends considerably on this l, research efforts have been estimating the optimal value for different scenarios, including 2D grids with multiple solutions. However, specifically
two previous works have used different stopping conditions for the simulations. Here, firstly, we show that these stopping conditions are not interchangeable. After doing such a pending investigation to define the stopping condition properly, we analyze the impacts of multiple solutions on the final results achieved by
the technique, which is the main contribution of this work. In doing so, we demonstrate that the success probability is inversely proportional to the density of vertices marked as solutions and directly proportional to the relative distance between solutions. These relations presented here are guaranteed only for high
values of the input parameters because, from different points of view, we show that a disturbed transition range exists in the small values.

Impacts of Multiple Solutions on the Lackadaisical Quantum Walk Search Algorithm

Standard features used for Credit Scoring includes mainly registration and financial data from customers. However, exploring new features is of great interest for financial companies, since slight improvements in the person score directly impact the company revenue. In this work, we categorize features from open credit scoring datasets and compare them with the features found in a real company dataset. The company dataset contains unusual feature groups such as historical, geolocation, web behavior, and demographic data. We performed bivariate tests using the Kolmogorov-Smirnov metric and features to assess the performance of the particular feature groups. We also generated a score of good payer by using AdaBoost, Multilayer Perceptron, and XGBoost algorithms. Then, we analyzed the results with different metrics and compared them with the real company results. Our main finding was that these features added a small improvement to current datasets. We also identified the most promising feature groups and noticed that the tuned XGBoost performed better than the company solution in three out of four deployed metrics.


Impact of Unusual Features in Credit Scoring Problem

The political campaigns in Brazilian elections are mostly financed by public money. Every candidate has to provide detailed accountability reports to the legal authorities, which must be analyzed in a short time frame in search of eventual fraud or suspicious transactions. In this work we have compiled a real data set from 2016 Brazilian elections for all city councils in the São Paulo state and used it to propose a framework of data segmentation analysis and validation. An exploratory data analysis is performed to determine the features distribution and to define the required feature pre-processing tasks. A clustering analysis using DBSCAN method is applied to a subset of the original data, focused on segmenting the spending data regarding contracts with car fuel providers and detecting potential outliers. Three clusters were identified and a ridge regression model was used to evaluate the most important features on cluster definition. One cluster was related to candidates that received zero votes and the remaining two discriminated suppliers if they had or not contracts almost exclusively related to candidate spending on car fuel. The hyperparameters from the clustering analysis were validated using a bootstrap method and a null hypothesis of data set structure randomness was rejected using a Monte Carlo approach.

Spending Segmentation and Outlier Detection in Brazilian Elections

Probabilistic circuits are deep probabilistic models with neural-network-like semantics capable of accurately and efficiently answering probabilistic queries without sacrificing expressiveness. Probabilistic Sentential Decision Diagrams (PSDDs) are a subclass of probabilistic circuits able to embed logical constraints to the circuit's structure. In doing so, they obtain extra expressiveness with empirical optimal performance. Despite achieving competitive performance compared to other state-of-the-art competitors, there have been very few attempts at learning PSDDs from a combination of both data and knowledge in the form of logical formulae. Our work investigates sampling random PSDDs consistent with domain knowledge and evaluating against state-of-the-art probabilistic models. We propose a method of sampling that retains important structural constraints on the circuit's graph that guarantee query tractability. Finally, we show that these samples are able to achieve competitive performance even on larger domains

Learning Probabilistic Sentential Decision Diagrams by Sampling

In recent years, bitcoin has become a very attractive investment in financial industry, which is not controlled by governments, but is based on trust between transfers under the technology of block chain. Hence, forecasting future bitcoin cryptocurrency values is a problem that has attracted the attention of many researchers in the field, while proving to be a very challenging problem. This work presents an experimental analysis using LSTM and GRUs for forecasting bitcoin values in a minute-granulated time for the entire next day. To this end we also present our methodology for conducting the experiments. The final goal is to create the core of a financial prediction tool around the RNNs. In our experiments, we achieved interesting results such as a SMAPE of 0.0002, a RMSE of US$ 3.844 and a rRMSE of 0.0028 in a day where bitcoin rates vary from US$ 13.2K and US$ 14.6K, surpassing the results of SMAPE found in the literature and proposed limit of SMAPE smaller than 0.007 for forecasts.

An Experimental Analysis of Model Compression Techniques for Object Detection

Downloads

Next from BRACIS 2020

Dense Captioning using Abstract Meaning Representation

Similar lecture

Automatic Student Network Search for Knowledge Distillation

An Experimental Analysis of Model Compression Techniques for Object Detection

.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

Dense Captioning using Abstract Meaning Representation

Similar lecture

Automatic Student Network Search for Knowledge Distillation

Downloads