LatinX in AI (LXAI) Research at ICML 2021

Maria Luisa Santiago, Miguel Alonso Jr, Laura Montoya, William Berrios, Fiorela Manco Fernández, Diana Diaz, Vinicius Caridá, LOURDES RAMIREZ CERNA, Pedro Braga, Gabriel Ramos, Leonel Rozo, Walter Mayor, Vanessa Gilede, Dennis Hernando Núñez Fernández, Erick Mendez Guzman, Paola Cascante-Bonilla


Launched in January 2018, leaders from academia and industry in Artificial Intelligence, Education, Research, Engineering, and Social Impact banded together to create a group that would be focused on “Creating Opportunity for LatinX in AI.”

Chat is not available.

Timezone: »


Mon 8:00 a.m. - 8:15 a.m.
Opening Remarks (Remark)   
Miguel Alonso Jr
Mon 8:15 a.m. - 8:45 a.m.
Marynel Vázquez (Talk)   
Marynel Vázquez
Mon 8:45 a.m. - 8:55 a.m.
Marynel Vázquez Q&A (Q&A)
Marynel Vázquez
Mon 8:55 a.m. - 9:25 a.m.

Presentation: Computation of Discrete Flows Over Networks via Constrained Wasserstein Barycenters Sparse Data Generation for Particle-Based Simulation of Hadronic Jets in the LHC Image Inpainting Applied to Art: Completing Escher’s Print Gallery Anticipating faults by predicting non-linearity of environment variables with neural networks: a case study in semiconductor manufacturing OCDE: Odds Conditional Density Estimator Deep Neural Network Uncertainty Estimation with Stochastic Inputs for Robust Aerial Navigation Policies

Mon 8:55 a.m. - 9:00 a.m.

We study a Wasserstein attraction approach for solving dynamic mass transport problems over networks. In the transport problem over networks, we start with a distribution over the set of nodes that needs to be "transported" to a target distribution accounting for the network topology. We exploit the specific structure of the problem, characterized by the computation of implicit gradient steps, and formulate an approach based on discretized flows. As a result, our proposed algorithm relies on the iterative computation of constrained Wasserstein barycenters. We show how the proposed method finds approximate solutions to the network transport problem, taking into account the topology of the network, the capacity of the communication channels, and the capacity of the individual nodes.

Ferran Arque, Cesar Uribe
Mon 9:00 a.m. - 9:05 a.m.

We develop a generative neural network for the generation of sparse data in particle physics using a permutation-invariant and physics-informed loss function. The input dataset used in this study consists of the particle constituents of hadronic jets due to its sparsity and the possibility of evaluating the network's ability to accurately describe the particles and jets properties. A variational autoencoder composed of convolutional layers in the encoder and decoder is used as the generator. The loss function consists of a reconstruction error term and the Kullback-Leibler divergence between the output of the encoder and the latent vector variables. The permutation-invariant loss on the particles' properties is combined with two mean-squared error terms that measure the difference between input and output jets mass and transverse momentum, which improves the network's generation capability as it imposes physics constraints, allowing the model to learn the kinematics of the jets.

Breno Orzari, Thiago Tomei, Maurizio Pierini, Mary Touranakou, Javier Duarte, Raghav Kansal, Dimitrios Gunopulos, jean-roch vlimant
Mon 9:05 a.m. - 9:10 a.m.

This work presents the first stages of a research in inpainting suited for art reconstruction. We introduce M.C Escher’s Print Gallery lithography as a use-case example. This artwork presents a void on its center and additionally, it follows a challenging mathematical structure that needs to be preserved by the inpainting method. We present our work so far and our future line of research.

Lucia Cipolina Kun
Mon 9:10 a.m. - 9:15 a.m.

The present work proposes a neural network model capable of anticipating possible faults in a semiconductor manufacturing plant by predicting non-linearity spikes in sensor data. Early detection of significant variation can be crucial for identifying machinery degradation or issues in the process itself. We use non-linearity as it is not affected by regular process changes and autocorrelation, thus avoiding false-positives in the neural network caused by changes in demand and the presence of control systems. The developed model is able to predict up to 30min of future non-linearity with loss ≤ 0.5. Furthermore, the proposed model is flexible enough to present itself as a starting point for future work in the field of fault detection in other areas.

Mateus Melchiades, Gabriel Ramos, Lincoln Schreiber
Mon 9:15 a.m. - 9:20 a.m.

Conditional density estimation (CDE) models can be useful for many statistical applications, especially because the full conditional density is estimated instead of point estimates, revealing more information about the uncertainty of random variables. In this paper, we propose a new methodology called Odds Conditional Density Estimator (OCDE) to estimate conditional densities in a supervised learning scheme. The main idea is that it is very difficult to estimate p{x,y} and p{x} in order to have the conditional density p_{y|x}, but by introducing an instrumental distribution, we transform the CDE problem into a problem of odds estimation, or similarly, training a binary probabilistic classifier. We demonstrate how OCDE works using simulated data and then test its performance against other known state-of-the-art CDE methods in real data. Overall, OCDE is competitive compared with these methods in real datasets.

Alex Aki Okuno, Felipe Polo
Mon 9:20 a.m. - 9:25 a.m.

It is well-known in the literature that uncertainty estimation methods are required in robotic autonomous systems that include deep learning (DL) components to assess the confidence in the outputs. However, to successfully deploy DL components in autonomous systems, they should also handle uncertainty at the input rather than only at the output. In this paper, we present a method to account for uncertainty at the input of Bayesian Deep Learning control policies for Aerial Navigation. Our experiments show that the proposed method improves the robustness of the navigation policy in out-of-distribution (OoD) scenarios.

Fabio Arnez Yagualca, Huascar Espinoza, François Terrier
Mon 9:25 a.m. - 9:35 a.m.
Spotlights 1 Q&A (Q&A)   
Leonel Rozo
Mon 9:35 a.m. - 9:45 a.m.
Mon 9:45 a.m. - 10:15 a.m.
Celia Cintas (Talk)   
Celia Cintas
Mon 10:15 a.m. - 10:25 a.m.
Celia Cintas Q&A (Q&A)
Celia Cintas
Mon 10:25 a.m. - 10:55 a.m.

Presentation: A multiple strategy for plant species identification using images of leaf texture Ceramic Cracks Segmentation with Deep Learning Towards Explainable Deep Reinforcement Learning for Traffic Signal Control Spatial Attention Adapted to a LSTM Architecture with Frame Selection for Human Action Recognition in Videos Minimum-Delay Adaptation in Non-Stationary Reinforcement Learning via Online High-Confidence Change-Point Detection Generalized linear tree: a flexible algorithm for predicting continuous variables

Mon 10:25 a.m. - 10:30 a.m.

In our planet there are thousands of plant species, being important to catalog these to help in the biodiversity preservation. However, identifying various plant species is not an easy task, even for specialists. Methods of computer vision for identifying plant species are interesting solutions for these difficulties. This work aims to analyze the efficiency of texture feature extraction methods applied in the identification of plant species by means of images of its leaves. For this, different texture descriptors were applied in three different databases. The obtained results indicate that local phase quantization (LPQ)-based methods achieve great efficiency and robustness. Additionally, the combination of LPQ-based methods with a segmentation based fractal texture analysis (SFTA) has increased the correct classification rate in all databases.

Igor Luidji Turra, Sérgio Francisco Silva, Douglas Cordeiro, Núbia Da Silva
Mon 10:30 a.m. - 10:35 a.m.

Cracks are pathologies whose appearance in ceramic tiles can cause various types of scratches due to the coating system losing water tightness and impermeability functions. Manual inspection is the most common method for this problem. However, it depends on the knowledge and experience of those who perform the analysis and demands a long time to map the entire area and high cost. These inspections require special equipment when they are at high altitudes, and the integrity of the inspector is at risk. Thus, there exists a need for automated optical inspection to find faults in ceramic tiles. This work focuses on the segmentation of cracks in ceramic images. We propose an architecture for segmenting cracks in facades with Deep Learning that includes a pre-processing step. We also propose the Ceramic Crack Database, a set of images to segment defects in ceramic tiles. The results show that the proposed architecture for ceramic crack segmentation achieves promising performance.

Gerivan Junior, Janderson Ferreira, Cristian Millán, Ramiro Ruiz, Alberto Junior, Bruno Fernandes
Mon 10:35 a.m. - 10:40 a.m.

Deep reinforcement learning has shown potential for traffic signal control. However, the lack of explainability has limited its use in real-world conditions. In this work, we present a Deep Q-learning approach, with the SHAP framework, able to explain its policy. Our approach can explain the impact of features on each action, which promotes the understanding of how the agent behaves in the face of different traffic conditions. Furthermore, our approach improved travel time, waiting time, and speed by 21.49%, 27.97%, 20.87%, compared to fixed-time traffic signal controllers.

Lincoln Schreiber, Gabriel Ramos, Ana Lucia Cetertich Bazzan
Mon 10:40 a.m. - 10:45 a.m.
Action recognition in videos is currently a topic of interest in the area of computer vision, due to potential applications such as: multimedia indexing, surveillance in public spaces, among others. In this work we propose an attention mechanism adapted to a CNN--LSTM base architecture. To carry out the training and testing phases, we used the HMDB-51 and UCF-101 datasets. We evaluate the performance of our system using accuracy as the evaluation metric, obtaining $57.3\%$ and $90.4\% $ for HMDB-51 and UCF-101 respectively.
ciorozco Orozco, María Elena Buemi, Julio Jacobo Berlles
Mon 10:45 a.m. - 10:50 a.m.

Non-stationary environments are challenging for reinforcement learning algorithms. If the state transition and/or reward functions change based on latent factors, the agent is effectively tasked with optimizing a behavior that maximizes performance over a possibly infinite random sequence of Markov Decision Processes (MDPs), each of which drawn from some unknown distribution. We call each such MDP a context. We introduce an algorithm that analyzes a possibly infinite stream of data and computes, in real-time, high-confidence change-point detection statistics that reflect whether novel, specialized policies need to be created and deployed to tackle novel contexts, or whether previously-optimized ones might be reused. We show that (i) this algorithm minimizes the delay until unforeseen changes to a context are detected, thereby allowing for rapid responses; and (ii) it bounds the rate of false alarm, which is important in order to minimize regret.

Lucas Alegre, Ana Lucia Cetertich Bazzan , Bruno C. da Silva
Mon 10:50 a.m. - 10:55 a.m.

Tree-based models are popular among regression methods to predict continuous variables. Also, Generalized Linear Models (GLMs) are pretty standard in many statistical applications and provide a generalization to many of the most commonly applied statistical procedures. However, in most regression tree methods, there is only one theoretical model associated for prediction in the final nodes, like multiple linear regression, logistic regressions, polynomial models, Poisson models, among others. We, therefore, propose a new tree method in which we estimate a GLM in each leaf node of the estimated tree including variable selection, new hyperparameters optimization, and tree pruning. Our method (GLT) has shown to be competitive compared to other well-known regression methods in real datasets, with the advantages and estimation flexibility provided by GLMs.

Alberto Rodrigues Ferreira, Alex Aki Okuno
Mon 10:55 a.m. - 11:05 a.m.
Spotlight 2 Q&A (Q&A)   
Cesar Uribe
Mon 11:05 a.m. - 12:05 p.m.
Mentoring Session 1 (Mentoring)
Pedro Braga, Diana Diaz, Vinicius Caridá
Mon 12:05 p.m. - 1:05 p.m.
Long (lunch + sponsors) break (Break)  link »
Mon 1:05 p.m. - 1:35 p.m.
Estevam R. Hruschka Jr (Talk)   
Estevam R. Hruschka Jr
Mon 1:35 p.m. - 1:45 p.m.
Estevam R. Hruschka Jr Q&A (Talk)
Estevam R. Hruschka Jr
Mon 1:45 p.m. - 2:15 p.m.
Spotlights 3 (Spotlight)
Mon 1:45 p.m. - 1:50 p.m.

This paper focuses on creating a system for recognizing the correct use of a mask through computer vision techniques. Research was carried out with aims of establishing the criteria for the creation of custom datasets, which were used to train, validate and test a pair of deep learning models, Mask-net and I-Mask-net. Both were designed with similar architectures, making use of Transfer Learning Techniques. The results given by training showed that the fine tuning carried out was adequate, while the tests carried out showed that the models have an acceptable level of accuracy, reaching 85.47% for Mask-net and 85.96% for I-Mask-net, additionally supported by the obtained precision, recall and F1-Score calculations.

Alexander Kalen, Alberto Landi, Nicolas Araque, Alejandro Marcano
Mon 1:50 p.m. - 1:55 p.m.

Social networks play a fundamental role in propagation of information and news. Characterizing the content of the messages becomes vital for tasks like fake news detection or personalized message recommendation. However, Twitter posts are short and often less coherent than other text documents, which makes it challenging to apply text mining algorithms efficiently. We propose a new pooling scheme for topic modeling in Twitter, which groups tweets whose authors belong to the same community on the retweet network into a single document. Our findings contribute to an improved methodology for identifying the latent topics in a Twitter dataset, without modifying the basic machinery of a topic decomposition model. In particular, we used Latent Dirichlet Allocation (LDA) and empirically showed that this novel method achieves better results than previous pooling methods in terms of cluster quality, document retrieval tasks, supervised machine learning classification and overall run time.

Federico Albanese
Mon 1:55 p.m. - 2:00 p.m.

Student's grade retention is a key issue faced by many education systems, especially those in developing countries. In this paper, we seek to gauge the relevance of students' personality traits in predicting grade retention in Brazil. For that, we used data collected in 2012 and 2017, in the city of Sertãozinho, countryside of the state of São Paulo, Brazil. The surveys taken in Sertãozinho included several socioeconomic questions, standardized tests, and a personality test. Moreover, students were in grades 4, 5, and 6 in 2012. Our approach was based on training machine learning models on the surveys’ data to predict grade retention between 2012 and 2017 using information from 2012 or before, and then using some strategies to quantify personality traits' predictive power. We concluded that, besides proving to be fairly better than a random classifier when isolated, personality traits contribute to prediction even when using socioeconomic variables and standardized tests results.

Lucka G Gianvechio, Carmen Toledo, Jonathan Ferreira, Felipe Polo, Renato Vicente
Mon 2:00 p.m. - 2:05 p.m.

Aspect-Based Sentiment Analysis (ABSA) tasks aim to identify consumers' opinions about different aspects of products or services. BERT-based language models have been used successfully in applications that require a deep understanding of the language, such as sentiment analysis. This paper investigates the use of disentangled learning to improve BERT-based textual representations in ABSA tasks. Motivated by the success of disentangled representation learning in the field of computer vision, which aims to obtain explanatory factors of the data representations, we explored the recent DeBERTa model (Decoding-enhanced BERT with Disentangled Attention) to disentangle the syntactic and semantics features from a BERT architecture. Experimental results show that incorporating disentangled attention and a simple fine-tuning strategy for downstream tasks outperforms state-of-the-art models in ABSA's benchmark datasets.

Ricardo Marcacini
Mon 2:05 p.m. - 2:10 p.m.

Early identification of patients with COVID-19 is essential to enable adequate treatment and to reduce the burden on the health system. The gold standard for COVID-19 detection is the use of RT-PCR tests. However, due to the high demand for tests, these can take days or even weeks in some regions of Brazil. Thus, an alternative for the detection of COVID-19 is the analysis of Chest X-rays (CXR). This paper proposes the evaluation of convolutional neural networks to identify pneumonia due to COVID-19 in CXR. The proposed methodology consists of an evaluation of six convolutional architectures pre-trained with the ImageNet dataset: InceptionResNetV2, InceptionV3, MovileNetV2, ResNet50, VGG16, and Xception. The obtained results for our methodology demonstrate that the Xception architecture presented a superior performance in the classification of CXR, with an Accuracy of 85.64%, Sensitivity of 85.71%, Specificity of 85.65%, F1-score of 85.49%, and an AUC of 0.9648.

Felipe Zeiser, Cristiano André da Costa, Gabriel Ramos
Mon 2:10 p.m. - 2:15 p.m.

Vehicle counting systems can help with vehicle analysis and traffic incident detection. Unfortunately, most existing methods require some level of human input to identify the Region of interest (ROI), movements of interest, or to establish a reference point or line to count vehicles from traffic cameras. This work introduces a method to count vehicles from traffic videos that automatically identifies the ROI for the camera, as well as the driving trajectories of the vehicles. This makes the method feasible to use with Pan-Tilt-Zoom cameras, which are frequently used in developing countries. Preliminary results indicate that the proposed method achieves an average intersection over the union of 57.05% for the ROI and a mean absolute error of just 17.44% at counting vehicles of the traffic video cameras tested.

Malolan Vasu, Christian Lopez
Mon 2:15 p.m. - 2:25 p.m.
Spotlights 3 Q&A (Q&A)   
Juan Banda
Mon 2:25 p.m. - 2:35 p.m.
Mon 2:35 p.m. - 3:05 p.m.
Prof. Ana L.C. Bazzan (Talk)   
Ana Lucia Cetertich Bazzan
Mon 3:05 p.m. - 3:15 p.m.
Prof. Ana L.C. Bazzan Q&A (Q&A)
Ana Lucia Cetertich Bazzan
Mon 3:15 p.m. - 3:45 p.m.
Spotlights 4 (Spotlights)
Mon 3:15 p.m. - 3:20 p.m.
[ Visit Poster at Spot A3 in Virtual World ]   

Although Machine Learning algorithms are solving tasks of ever-increasing complexity, gathering data and building training sets remains an error-prone, costly, and difficult problem. However, reusing knowledge from related previously-solved tasks enables reducing the amount of data required to learn a new task. We here propose a method for reusing a tree-based model learned in a source task with abundant data in a target task with scarce data. We perform an empirical evaluation showing that our method is useful, especially in scenarios where the labels are unavailable in the target task.

Leno Silva, Renato Vicente
Mon 3:20 p.m. - 3:25 p.m.

Although Machine Learning algorithms are solving tasks of ever-increasing complexity, gathering data and building training sets remains an error prone, costly, and difficult problem. However, reusing knowledge from related previously-solved tasks enables reducing the amount of data required to learn a new task. We here propose a method for learning a mapping model that maps data from a source task with labeled data to a related target task with only unlabeled data. We perform an empirical evaluation showing that our method achieves performance comparable to a model learned directly in the target task.

Leno Silva, Ruben Glatt, Renato Vicente
Mon 3:25 p.m. - 3:30 p.m.

We study the problem of minimizing a convex function over probability measures supported in a graph. We build upon the recent formulation of optimal transport over discrete domains to propose a method that generates a sequence that provably converges to a minimum of the objective function and smoothly transports mass over the edges of the graph. Moreover, we identify novel relation between Riemannian gradient flows and perturbed best response protocols that provide sufficient conditions for the convergence of the proposed algorithm. Numerical results show practical advantages over existing approaches with respect to the implementability and convergence rates.

Gilberto Díaz-García, Cesar Uribe, Nicanor Quijano
Mon 3:30 p.m. - 3:35 p.m.

We propose an online adaptive synchronization method for leader-follower networks of heterogeneous agents. Synchronization is achieved using a distributed Model Reference Adaptive Control (DMRAC-RL) that enables the improved performance of Reinforcement Learning (RL)-trained policies on a reference model. The leader observes the performance of the reference model, and the followers observe the states and actions of the agents they are connected to, but not the reference model. Notably, both the leader and followers models might differ from the reference model the RL-control policy was trained. DMRAC-RL uses an internal loop that adjusts the learned policy for the agents in the form of an augmented input to solve the distributed control problem. Numerical examples of the synchronization of a network of inverted pendulums support our theoretical findings.

Miguel F. Arevalo-Castiblanco, Cesar Uribe, Eduardo Mojica-Nava
Mon 3:35 p.m. - 3:40 p.m.

We study the robustness properties of computationally-aware dual-based distributed optimization algorithm over networks. Contrary to existing literature, we follow a hybrid dynamical systems approach to analyze the stability properties of the distributed Nesterov's ODE when explicitly taking into account the computational resources and time required by a dual first-order oracle to generates an approximate gradient. We show that in such scenario, the distributed Nesterov's ODE is unstable in the Lyapunov sense, i.e., there exist an arbitrarily bounded perturbation function for which the inexact oracle drives the system unstable. Moreover, we propose modified dynamics that are provable stable and robust, and which provably minimizes smooth and strongly convex functions.

Daniel Ochoa, Jorge Poveda, Cesar Uribe
Mon 3:40 p.m. - 3:45 p.m.

Learning to process long-range dependencies has been a challenge for recurrent neural networks. Despite improvements achieved by long short-term memory (LSTMs), its gating mechanism results in exponential decay of information, limiting their capacity of capturing long-range dependencies. In this work, we present a power law forget gate, which instead has a slower rate of information decay. We propose a power law-based LSTM (pLSTM) based on the LSTM but with a power law forget gate. We test empirically the pLSTM on the copy task, sentiment classification, and sequential MNIST, all with long-range dependency tasks. The pLSTM solves these tasks outperforming an LSTM, specially for long-range dependencies. Further, the pLSTM learns sparser and more robust representations.

Hsiang-Yun Chien, Javier Turek, Nicole Beckage, Vy Vo, Christopher Honey, Theodore Willke
Mon 3:45 p.m. - 3:55 p.m.
Spotlights 4 Q&A (Q&A)   
Gabriel Ramos
Mon 3:55 p.m. - 4:55 p.m.
Mentoring II (Mentoring)   
Pedro Braga, Diana Diaz, Vinicius Caridá, LOURDES RAMIREZ CERNA, Paola Cascante-Bonilla
Mon 4:55 p.m. - 5:00 p.m.
Closing Remarks (Remarks)
Gabriel Ramos
Mon 5:00 p.m. -
Social Gathering  link »