Message Passing Adaptive Resonance Theory for Online Active Semi-supervised Learning

Abstract

Active learning is widely used to reduce labeling effort and training time by repeatedly querying only the most beneficial samples from the unlabeled data. In real-world problems where data cannot be stored indefinitely due to limited storage or privacy issues, the query selection and the model update should be performed as soon as a new data sample is observed. Various online active learning methods have been studied to deal with these challenges; however, there are difficulties in selecting representative query samples and updating the model efficiently. In this study, we propose Message Passing Adaptive Resonance Theory (MPART) for online active semi-supervised learning. The proposed model learns the distribution and topology of the input data online. It then infers the class of unlabeled data and selects informative and representative samples through message passing between nodes on the topological graph. MPART queries the beneficial samples on-the-fly in stream-based selective sampling scenarios, and continuously improve the classification model using both labeled and unlabeled data. We evaluate our model on visual (MNIST, SVHN, CIFAR-10) and audio (NSynth) datasets with comparable query selection strategies and frequencies, showing that MPART significantly outperforms the competitive models in online active learning environments.

CitySurfaces: City-Scale Semantic Segmentation of Sidewalk Materials

“…The conventional approach in active learning is to select new samples (query) every iteration, which can work in cases where the cost of annotation is not high or in experimental studies that work with already annotated images to advance the field and develop new query algorithms, as is the case with most of the already published works in active learning for semantic segmentation, where they use datasets such as Cityscapes (Cordts et al, 2016) or ADE20k (Zhou et al, 2017). However, since no annotated dataset exists for sidewalk materials, we have to annotate every new sample we choose during the training process, and it is impractical to annotate a new sample for every iteration (Kim et al, 2020). To overcome this, we adopt a multistage framework and annotate a new sample at the end of each stage, where each stage consists of ten epochs.…”