Methods,Applications,and Implementation Of Weakly Supervised Learning On Zero/Few-Shot Scenarios

With the increases of the amount of data and information in modern society,a series of artificial intelligence technologies have made important breakthroughs.Particularly,supervised learning models have exceeded human in many laboratory scenarios.However,training an effective supervised learning model usually needs to collect hundreds of manually labeled training samples for each target category.In practice,due to the long tail effect of real data category distribution and the high cost of manual labeling,it is often impossible to establish an ideal supervised training dataset for the model,resulting in a variety of zero-sample and few-sample scenarios,such as the lack of samples or labels for certain categories.Aiming at the widespread and important insufficient data problem,researchers in the machine learning field have proposed some solutions that have less dependence on the labeled dataset,which are collectively termed as weakly supervised learning.Its deep-seated principle is to fully tap the characteristics of existing datasets,so that the establishment of intelligent model is more in line with human learning mechanism,so as to have real machine intelligence step by step.Therefore,the research on weakly supervised learning for the zero-sample or few-sample scenarios is becoming a hot issue in many modeling fields.Based on the above background,this paper takes weakly supervised learning as the core task,starts from the basic few-sample scenario and few-label scenario,deduces to the zerosample scenario,and then to the final any-sample scenario.We gradually reveal and solve some key problems in the establishment of weakly supervised learning model,so as to expand the application of data-driven model in various weakly supervised scenarios.In order to present the realistic existence and practical value of weakly supervised learning problems,such as fewsample scenario,few-label scenario,zero-sample scenario,and any-sample scenario,we validate the proposed weakly supervised models with four typical applications: handwritten digit recognition,virtual measurement,industrial fault diagnosis,and outdoor scene detection.At the same time,we design and open source a generative model repository for weakly supervised learning based on Python language,so that the generative model designed in this paper and a series of representative models can be reproduced with only one command line.The main contributions of this paper are divided into five points,which are summarized as follows:1)For the few-sample scenario in weakly supervised learning,a broad network gradient boosting model with triple incremental learning ability is designed.The new model is constructed by combining multiple broad networks with the additive model of gradient boosting machine,which imitating the deep structure of convolutional neural network.It addresses the problem that a model established with few samples has to be retrained from the beginning for learning new features,new samples,and new categories.In the handwritten digit recognition experiment,it reduces the modeling time by 40%～60%.2)For the few-label scenario in weakly supervised learning,a simple but effective semisupervised adversarial smoothing regularization loss is designed.The regularization term measures the local smoothness of the model’s prediction around each input sample,and addresses the robustness problem of modeling with a large number of unlabeled samples,by minimizing the divergence of model prediction for noisy and clean samples.At the same time,we design a semisupervised triple regression framework to further use unlabeled samples with pseudo labels,and then construct an adversarial smoothing triple regression model for robust semisupervised virtual measurement,which reduces the prediction error by 7%～10% and the noise error by 25%.3)For the zero-sample scenario in weakly supervised learning,a zero-sample model based on semantic description attribute transfer is designed.Under the data-driven framework,the model attempts to use the manually defined description instead of the sample to determine the category,which addresses the modeling problem for categories which have no training samples and labels.Taking the industrial fault diagnosis task as an example,through pretraining and knowledge transfer from existing faults,the zero-sample model could diagnose target faults online based on the identified fault description without any data-based model training,and achieves the accuracy by a model established with 200 ～400 samples.4)For the any-sample scenario with zero-sample and few-sample problems in weakly supervised learning,a semantic refinement generative adversarial network with the ability of eliminating the bias of generator transfer is designed.The model uses multihead semantic representation technique and hierarchical semantic alignment technique to refine the semantic description,which addresses the bias problem in the feature generator transfer from seen categories to unseen categories,so as to achieve the bias-eliminated condition for the disjoint-class generator transfer.Taking the outdoor scene detection task as an example,in comparison with some conventional generative models,the accuracy of the new model is improved by 5%～8% in the zero-sample test and by 5%～10% in the few-sample test.5)For weakly supervised learning,a generative model repository is designed based on Python language and is made publicly available.It provides models,features,parameters,and experimental settings.The generative model designed in this paper and some typical generative models for zero-sample or few-sample problems can be reproduced with only one command line.At the same time,based on the outdoor scene detection task,we significantly improve generative models’ performance by simply modifying the visual and semantic features,revealing the importance of the visual and semantic embeddings for generative model and weakly supervised learning.