Adversarial Sampling for Active Learning (ASAL) Adversarial Sample Generation using GANs Discriminator 的作用是确保generator 生成的样本与真实样本相差无几。作者的想法是从Generator的输出中进行Uncertainty Sampling,这样uncertainty sampling的两种策略可以表示为: 以及 但是作者同时指出,直接生成样本有两个问题: 需要人...
Adversarial Sampling for Active Learningdoi:10.1109/WACV45572.2020.9093556Christoph MayerRadu TimofteIEEEWorkshop on Applications of Computer Vision
1、传统的uncertainty sampling active learning效率太低了,需要频繁轮询整个unlabeled dataset。 2、本文的一个有趣的intution是,基于一个labeled example生成的adversarial example附近的unlabeled example是高质量example,因为adversarial example一般都离decision boundary很近,这个距离正是uncertainty,距离越小,uncertainty越大...
28. Model-Agnostic Boundary-Adversarial Sampling for Test-Time Generalization in Few-Shot learning 会议:ECCV 2020. 作者:Jaekyeom Kim, Hyoungseok Kim, Gunhee Kim 链接:https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123460579.pdf 29. Multitask Learning Strengthens Adversarial Robustness ...
This work proposes an active learning method based on deep AAE-SDR neural network for high-dimensional reliability analysis. A sampling strategy based on the latent space is proposed to iteratively select misclassified samples that contribute significantly to enhancing the accuracy of the AAE-SDR neura...
Variational Autoencoders: Autoencoders consist of a pair of connected networks: the encoder and the decoder. Variational autoencoders are distinguished by having a continuous latent space. This allows for easy sampling and interpolation, which is particularly useful when VAEs are used to explore va...
(e.g., different rotation, scale, and translation). We compute adversarial examples that are robust to image transformations by sampling random geometric transformations applied to the original image at each step of the perturbation algorithm (rotationθ ~ U(0,π/6), scalesx, sy ~...
We describe a pool-based semi-supervised active learning algorithm that implicitly learns this sampling mechanism in an adversarial manner. Unlike conventional active learning algorithms, our approach is \textit{task agnostic}, i.e., it does not depend on the performance of the task for which we...
现有的方法主要是UCB [1], Thompson Sampling (TS) [2] 这种类型的方法,它们用uncertainty来描述一个item的潜在收益。(因此就是将长期收益 & instant reward(实时点击率)的折中 换成了 uncertainty 和 instant reward的收益的折中。 但是,现有方法并没有去思考一个item的被曝光以后到得到用户反馈这个过程对模型的...
Parametric models for generating images has been explored extensively (for example on MNIST digits or for texture synthesis (Portilla & Simoncelli, 2000)). However, generating natural images of the real world have had not much success until recently. A variational sampling approach to generating ima...