Scalable teacher forcing network for semi-supervised large scale data streams

An explainable semi-supervised self-organizing fuzzy inference system for streaming data classification

“…As aforementioned, S 3 OFIS+ uses SOFIS+ [41] as its implementation basis and exploits the idea of "pseudo labelling" [32] to perform self-training from unlabelled streaming data on a chunk-by-chunk basis with the aim of constructing a stronger prediction model. As aforementioned, this study considers the highly challenging infinite delay problems [40], [43]. In such scenarios, only a reduced set of labelled data is available during the warming up stage [43].…”

“…To evaluate the performance of the proposed S 3 OFIS+ system, numerical examples based on a wide range of benchmark problems from UCI Machine Learning Repository 1 , Keel Dataset Repository 2 and Scikit-Multiflow 3 are presented. As aforementioned, all numerical examples in this paper are performed in infinite delay scenarios [40], [43]. The datasets used for experimental studies are as follows.…”

“…STHP can initialize its multi-layered prototype-based structure with a small set of labelled data from scratch and then continuously self-develop with new unlabelled samples. A weakly supervised scalable teacher forcing network for large-scale data streams (WeScatterNet) is presented in [40]. WeScatterNet features an ensemble framework with its base learners dynamically added and pruned via drift detection.…”

“…In this paper, a novel self-training EIS named semi-supervised self-organizing fuzzy inference system (S 3 OFIS+) is proposed for semi-supervised learning from streaming data in indefinite delay environments [40]. The proposed S 3 OFIS+ uses the simplified self-organizing fuzzy inference system (SOFIS+) [41] as its implementation basis.…”

VGFlow: Visibility guided Flow Network for Human Reposing

“…The ground truth for the visibility map V ismap gt , was obtained by fitting densepose [8] on I s , I t and then matching the acquired U V coordinates to generate the visible and invisible mask. We further use teacher forcing technique [24] for training F lowV is, in which the ground truth VisMap is used with 50% probability for the warping losses. The losses guiding the flow module can be summarized as follows.…”