Meta-learning considers the problem of learning an efficient learning process that can leverage its past experience to accurately solve new tasks. However, the efficacy of meta-learning crucially depends on the distribution of tasks available for training, and this is often assumed to be known a priori or constructed from limited supervised datasets. In this work, we aim to provide task distributions for meta-learning by considering self-supervised tasks automatically proposed from unlabeled text, to enable large-scale meta-learning in NLP. We design multiple distributions of self-supervised tasks by considering important aspects of task diversity, difficulty, type, domain, and curriculum, and investigate how they affect meta-learning performance. Our analysis shows that all these factors meaningfully alter the task distribution, some inducing significant improvements in downstream few-shot accuracy of the metalearned models. Empirically, results on 20 downstream tasks show significant improvements in few-shot learning -adding up to +4.2% absolute accuracy (on average) to the previous unsupervised meta-learning method, and perform comparably to supervised methods on the FewRel 2.0 benchmark.
Meta-learning considers the problem of learning an efficient learning process that can leverage its past experience to accurately solve new tasks. However, the efficacy of meta-learning crucially depends on the distribution of tasks available for training, and this is often assumed to be known a priori or constructed from limited supervised datasets. In this work, we aim to provide task distributions for meta-learning by considering self-supervised tasks automatically proposed from unlabeled text, to enable large-scale meta-learning in NLP. We design multiple distributions of self-supervised tasks by considering important aspects of task diversity, difficulty, type, domain, and curriculum, and investigate how they affect meta-learning performance. Our analysis shows that all these factors meaningfully alter the task distribution, some inducing significant improvements in downstream few-shot accuracy of the metalearned models. Empirically, results on 20 downstream tasks show significant improvements in few-shot learning -adding up to +4.2% absolute accuracy (on average) to the previous unsupervised meta-learning method, and perform comparably to supervised methods on the FewRel 2.0 benchmark.
For years, Single Image Super Resolution (SISR) has been an interesting and ill-posed problem in computer vision. The traditional super-resolution (SR) imaging approaches in- volve interpolation, reconstruction, and learning-based methods. Interpolation methods are fast and uncomplicated to compute, but they are not so accurate and reliable. Reconstruction-based methods are better compared with interpolation methods, but they are time-consuming and the quality degrades as the scaling increases. Even though learning-based methods like Markov random chains are far better than all the previous ones, they are unable to match the performance of deep learning models for SISR. This study examines the Residual Dense Networks architec- ture proposed by Yhang et al. [17] and analyzes the importance of its components. By leveraging hierarchical features from original low-resolution (LR) images, this architecture achieves superior performance, with a network structure comprising four main blocks, including the residual dense block (RDB) as the core. Through investigations of each block and analyses using various loss metrics, the study evaluates the effectiveness of the architecture and compares it to other state-of-the-art models that differ in both architecture and components.
Object occlusion is one of the most indispensable and challenging problems in computer vision. While convolutional neural networks (CNNs) have proven to be effective for regular image classification, they struggle with images that have partial occlusions. Partial occlusion is a scenario where an object is partially occluded by another object or space. This problem, when solved, holds tremendous potential to facilitate various scenarios. We, in particular, are interested in the autonomous driving scenario and its implications. Autonomous vehicle research is one of the hot topics of this decade. There are ample situations of partial occlusions of a driving sign, a person, or other objects at different angles. The importance of this problem extends beyond autonomous driving and could be exploited for video analytics in traffic data to handle crimes, anticipate income levels of various groups, and more. In this paper, we introduce our own synthetically created dataset by utilizing the Stanford Car Dataset and adding occlusions of various sizes and natures to it. then conduct a comprehensive analysis using various state-of-the-art models and study the effect of varying occlusion proportions and nature on the models' performance. This research provides insights into which models are more robust to partial occlusions and how they perform when trained with occluded and unoccluded images, which can aid in developing more reliable autonomous driving systems and other computer vision applications.
The task of predicting the publication period of text documents, such as news articles, is an important but less studied problem in the field of natural language processing. Predicting the year of a news article can be useful in various contexts, such as historical research, sentiment analysis, and media monitoring. In this work, we investigate the problem of predicting the publication period of a text document, specifically a news article, based on its textual content. In order to do so, we created our own extensive labeled dataset of over 350,000 news articles published by The New York Times over six decades. In our approach, we use a pretrained BERT model fine-tuned for the task of text classification, specifically for time period prediction.This model exceeds our expectations and provides some very impressive results in terms of accurately classifying news articles into their respective publication decades. The results beat the performance of the baseline model for this relatively unexplored task of time prediction from text.
The travel time studies are one of the most important measures used for evaluating the performance of road networks. The Global Positioning System (GPS) is a space-based system that provides position and time information in all weather conditions. GPS data could be used to obtain the values of traffic control delay, vehicle queue, average travel time and vehicle acceleration and deceleration at intersections.The task of estimation of delay becomes complex if it is performed for intersections carrying heterogeneous traffic and that to for over saturated conditions. Most of the urban signalized intersections are manually controlled during peak hours. GPS device fitted in a vehicle was run repeatedly during morning peak period and the period during which vehicles were allowed to cross the intersection was recorded with video graphic camera. The attempt to identify the control delay with the GPS data from the test vehicle while crossing manually operated major intersection is presented in this paper.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.