Destaques
Comunidades no Repositório
Selecione uma comunidade para navegar por suas coleções
Submissões Recentes
Exploration and Rescue of Shipwreck Survivors using Reinforcement Learning-Empowered Drone Swarms
(2023) Abreu, Leonardo D. M. de; Carrete, Luis F. S.; Castanares, Manuel; Damiani, Enrico F.; Brancalion, Jose Fernando B.; Barth, Fabrício J.
The goal of this project is to create a reinforcement learning algorithm that locates shipwrecked individuals using a swarm of drones. A simulated environment was developed to train and visualize the outcome of the trained algorithm considering the ocean’s dynamic circumstances. This project does not discuss image recognition of shipwrecked people, since the true focus of this project is to optimize the search routine of a drone to find the target in the most efficient way possible. The implemented Reinforce algorithm takes into account a dynamic map of probabilities, representing the chances of a person being found, as well as the position of other agents. Outcomes include an open-source Python package for the environment and the implementation of the reinforcement learning algorithm.
The algorithm demonstrates superiority over the predefined approach, proving the advantages of reinforcement learning in efficiency and effectiveness.
Unsupervised Improvement of Audio-Text Cross-Modal Representations
(2023) Wang, Zhepei; Subakan, Cem; Subramani, Krishna; Wu, Junkai; TIAGO FERNANDES TAVARES; FABIO JOSE AYRES; Smaragdis, Paris
Recent advances in using language models to obtain cross-modal audio-text representations have overcome the limitations of conventional training approaches that use predefined labels. This has allowed the community to make progress in tasks like zero-shot classification, which would otherwise not be possible. However, learning such representations requires a large amount of human-annotated audio-text pairs. In this paper, we study unsupervised approaches to improve the learning framework of such representations with unpaired text and audio. We explore domain-unspecific and domain-specific curation methods to create audio-text pairs that we use to further improve the model. We also show that when domain-specific curation is used in conjunction with a soft-labeled contrastive loss, we are able to obtain significant improvement in terms of zero-shot classification performance on downstream sound event classification or acoustic scene classification tasks.
Well-Connected Communities in Real-World and Synthetic Networks
(2023) Park, Minhyuk; Tabatabaee, Yasamin; Ramavarapu, Vikram; Liu, Baqiao; Pailodi, Vidya Kamath; Ramachandran, Rajiv; Korobskiy, Dmitriy; FABIO JOSE AYRES; Chacko, George; Warnow, Tandy
Integral to the problem of detecting communities through graph clustering is the expectation that they are "well connected". In this respect, we examine five different community detection approaches optimizing different criteria: the Leiden algorithm optimizing the Constant Potts Model, the Leiden algorithm optimizing modularity, Iterative K-Core Clustering (IKC), Infomap, and Markov Clustering (MCL). Surprisingly, all these methods produce, to varying extents, communities that fail even a mild requirement for well connectedness. To remediate clusters that are not well connected, we have developed the "Connectivity Modifier" (CM), which, at the cost of coverage, iteratively removes small edge cuts and re-clusters until all communities produced are well connected. Results from real-world and synthetic networks illustrate a tradeoff users make between well connected clusters and coverage, and raise questions about the "clusterability" of networks and models of community structure.
Using bundling to visualize multivariate urban mobility structure patterns in the São Paulo metropolitan area
(2021) Martins, Tallys G.; Lago, Nelson; Santana, Eduardo F. Z.; Telea, Alexandru; Kon, Fabio; Souza, Higor A. de
Internet-based technologies such as IoT, GPS-based systems, and cellular networks enable the collection of geolocated mobility data of millions of people in large metropolitan areas. In addition, large, public datasets are made available on the Internet by open government programs, providing ways for citizens, NGOs, scientists, and public managers to perform a multitude of data analysis with the goal of better understanding the city dynamics to provide means for evidence-based public policymaking. However, it is challenging to visualize huge amounts of data from mobility datasets. Plotting raw trajectories on a map often causes data occlusion, impairing the visual analysis. Displaying the multiple attributes that these trajectories come with is an even larger challenge. One approach to solve this problem is trail bundling, which groups motion trails that are spatially close in a simplified representation. In this paper, we augment a recent bundling technique to support multi-attribute trail datasets for the visual analysis of urban mobility. Our case study is based on the travel survey from the São Paulo Metropolitan Area, which is one of the most intense traffic areas in the world. The results show that bundling helps the identification and analysis of various mobility patterns for different data attributes, such as peak hours, social strata, and transportation modes.