MEST: Accurate and Fast Memory-Economic Sparse Training Framework on the Edge
Activation Modulation and Recalibration Scheme for Weakly Supervised Semantic Segmentation
Exploring the Equivalence of Siamese Self-Supervised Learning via A Unified Gradient Framework
A Mathematical Framework for Transformer Circuits
Towards Transferable Adversarial Attacks on Vision Transformers
Multi-Scale 2D Temporal Adjacent Networks for Moment Localization with Natural Language
ArXiv Weekly Radiostation:NLP、CV、ML 更多精选论文(附音频)
论文 1:ViR: the Vision Reservoir
作者:Xian Wei、Bin Wang、Mingsong Chen 等
论文链接:https://arxiv.org/pdf/2112.13545.pdf
摘要:近一年来,视觉 Transformer(ViT) 在图像任务上大放光芒,比如在图像分类、实例分割、目标检测分析和跟踪等任务上显示出了卓越的性能,展现出取代卷积神经网络的潜力。但仍有证据表明,在大规模数据集上应用多个 Transformer 层进行预训练时,ViT 往往存在两个方面的问题:一是计算量大,内存负担大;二是在小规模数据集上从零开始训练时存在过拟合问题。 为了解决这些问题,来自华东师范大学等机构的研究者们提出了一种新的图像分类方法,即 Vision Reservoir (ViR) 。通过将每个图像分割成一系列具有固定长度的 token,ViR 构建一个具有几乎完全连接拓扑的纯库,以替换 ViT 中的 Transformer 模块。为了提高网络性能,研究者还提出了两种深度 ViR 模型。
在 CIFAR100 数据集上执行 ViR 和 ViT 的时间消耗比较。
模型概述。 推荐:参数量下降 85%,性能全面超越 ViT,华人团队提出全新图像分类方法 ViR 论文 2:MEST: Accurate and Fast Memory-Economic Sparse Training Framework on the Edge
作者:Geng Yuan、Xiaolong Ma、Wei Niu 等
论文链接:https://arxiv.org/pdf/2110.14032.pdf
摘要:在剪枝技术被成功应用于神经网络的压缩和加速之后,稀疏训练在近年来受到了越来越多研究者的关注,即如何从零开始直接训练一个高质量的稀疏神经网络。稀疏训练旨在有效降低神经网络训练过程中的计算和存储开销,从而加速训练过程,为在资源有限的边缘设备上的神经网络训练提供了更多可能性。 多数现有稀疏训练方法着力于设计更好的稀疏训练算法来追求更高的网络稀疏度同时保持高准确率。然而,稀疏训练的 “真正精神”,即稀疏训练能否带来实际的训练加速以及计算和存储资源的节省,却往往被忽视了。为此,由美国东北大学王言治教授、林雪教授研究组与威廉玛丽学院任彬教授研究组共同提出了 MEST 稀疏训练框架,有望实现在边缘设备上的准确、快速以及内存经济的稀疏训练。
算法 1。
与 SOTA 方法的比较。 推荐:目前,该文章 [1] 已被 NeurIPS 2021 会议收录为 spotlight 论文。 论文 3:Activation Modulation and Recalibration Scheme for Weakly Supervised Semantic Segmentation
与 SOTA 方法的比较。 推荐:在图像级弱监督语义分割这项 CV 难题上,字节跳动做到了性能显著提升。 论文 4:Exploring the Equivalence of Siamese Self-Supervised Learning via A Unified Gradient Framework
不同攻击方法在 ViTs 上的攻击成功率结果对比。 推荐:无注意力 + PatchOut,复旦大学提出面向视觉 transformer 的迁移攻击方法。 论文 7:Multi-Scale 2D Temporal Adjacent Networks for Moment Localization with Natural Language
1. LINDA: Unsupervised Learning to Interpolate in Natural Language Processing. (from Kyunghyun Cho)2. Frequency-Aware Contrastive Learning for Neural Machine Translation. (from Tong Zhang)3. CUGE: A Chinese Language Understanding and Generation Evaluation Benchmark. (from Kun Zhou, Minlie Huang, Xiaodong He, Yang Liu)4. Measuring Attribution in Natural Language Generation Models. (from Michael Collins)5. Simple, Interpretable and Stable Method for Detecting Words with Usage Change across Corpora. (from Yoav Goldberg)6. Counterfactual Memorization in Neural Language Models. (from Chiyuan Zhang)7. Towards Personalized Answer Generation in E-Commerce via Multi-Perspective Preference Modeling. (from Bolin Ding)8. HeteroQA: Learning towards Question-and-Answering through Multiple Information Sources via Heterogeneous Graph Modeling. (from Dongyan Zhao)9. Parameter Differentiation based Multilingual Neural Machine Translation. (from Qian Wang)10. Visual Persuasion in COVID-19 Social Media Content: A Multi-Modal Characterization. (from Adriana Kovashka) 10 CV Papers音频:进度条00:00/21:22 本周 10 篇 CV 精选论文是: 1. HSPACE: Synthetic Parametric Humans Animated in Complex Environments. (from William T. Freeman, Rahul Sukthankar)2. Gendered Differences in Face Recognition Accuracy Explained by Hairstyles, Makeup, and Facial Morphology. (from Kevin W. Bowyer)3. TAGPerson: A Target-Aware Generation Pipeline for Person Re-identification. (from Kai Chen)4. Recursive Least-Squares Estimator-Aided Online Learning for Visual Tracking. (from Weiming Hu)5. Cross-Domain Empirical Risk Minimization for Unbiased Long-tailed Classification. (from Xian-Sheng Hua)6. Associative Adversarial Learning Based on Selective Attack. (from David Doermann)7. Extended Self-Critical Pipeline for Transforming Videos to Text (TRECVID-VTT Task 2021) -- Team: MMCUniAugsburg. (from Rainer Lienhart)8. Synchronized Audio-Visual Frames with Fractional Positional Encoding for Transformers in Video-to-Text Translation. (from Rainer Lienhart)9. Invertible Network for Unpaired Low-light Image Enhancement. (from Wangmeng Zuo)10. GuidedMix-Net: Semi-supervised Semantic Segmentation by Using Labeled Images as Reference. (from Ling Shao) 10 ML Papers音频:进度条00:00/20:34
本周 10 篇 ML 精选论文是:
1. Wasserstein Flow Meets Replicator Dynamics: A Mean-Field Analysis of Representation Learning in Actor-Critic. (from Michael I. Jordan)2. Can Reinforcement Learning Find Stackelberg-Nash Equilibria in General-Sum Markov Games with Myopic Followers?. (from Michael I. Jordan)3. Graph Few-shot Class-incremental Learning. (from Huan Liu)4. Sparsest Univariate Learning Models Under Lipschitz Constraint. (from Michael Unser)5. Adaptivity and Non-stationarity: Problem-dependent Dynamic Regret for Online Convex Optimization. (from Zhi-Hua Zhou)6. Differentially-Private Clustering of Easy Instances. (from Yishay Mansour)7. Disentanglement and Generalization Under Correlation Shifts. (from Richard Zemel)8. Efficient Performance Bounds for Primal-Dual Reinforcement Learning from Demonstrations. (from John Lygeros)9. Learning Optimization Proxies for Large-Scale Security-Constrained Economic Dispatch. (from Pascal Van Hentenryck)10. Exponential Family Model-Based Reinforcement Learning via Score Matching. (from Nathan Srebro)