Filters








45 Hits in 4.0 sec

Manifold Mixup: Better Representations by Interpolating Hidden States [article]

Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, Aaron Courville, David Lopez-Paz, Yoshua Bengio
2019 arXiv   pre-print
To address these issues, we propose Manifold Mixup, a simple regularizer that encourages neural networks to predict less confidently on interpolations of hidden representations.  ...  As a result, neural networks trained with Manifold Mixup learn class-representations with fewer directions of variance.  ...  This suggests that mixup acts as a useful regularization on the discriminator, which is even further improved by Manifold Mixup. (See Figure 8 for the full set of experimental results.)  ... 
arXiv:1806.05236v7 fatcat:66wh3msxd5ee7ga7xvdz3nncre

Charting the Right Manifold: Manifold Mixup for Few-shot Learning [article]

Puneet Mangla, Mayank Singh, Abhishek Sinha, Nupur Kumari, Vineeth N Balasubramanian, Balaji Krishnamurthy
2020 arXiv   pre-print
A recent regularization technique - Manifold Mixup focuses on learning a general-purpose representation, robust to small changes in the data distribution.  ...  We observe that regularizing the feature manifold, enriched via self-supervised techniques, with Manifold Mixup significantly improves few-shot learning performance.  ...  Manifold Mixup for Few-shot Learning Higher-layer representations in neural network classifiers have often been visualized as lying on a meaningful manifold, that provide the relevant geometry of data  ... 
arXiv:1907.12087v4 fatcat:a5v3zqnvyva3zar62g6as44u3a

Neighborhood Mixup Experience Replay: Local Convex Interpolation for Improved Sample Efficiency in Continuous Control Tasks [article]

Ryan Sander, Wilko Schwarting, Tim Seyde, Igor Gilitschenski, Sertac Karaman, Daniela Rus
2022 arXiv   pre-print
NMER preserves a locally linear approximation of the transition manifold by only applying Mixup between transitions with vicinal state-action features.  ...  Under NMER, a given transition's set of state action neighbors is dynamic and episode agnostic, in turn encouraging greater policy generalizability via inter-episode interpolation.  ...  Neighborhood Mixup as a heuristic to encourage on-manifold interpolation.  ... 
arXiv:2205.09117v1 fatcat:o5bvbw77u5dbjnhbzqre4joo5i

Few-shot Image Generation with Mixup-based Distance Learning [article]

Chaerin Kong, Jeesoo Kim, Donghoon Han, Nojun Kwak
2022 arXiv   pre-print
We propose mixup-based distance regularization on the feature space of both a generator and the counterpart discriminator that encourages the two players to reason not only about the scarce observed data  ...  Producing diverse and realistic images with generative models such as GANs typically requires large scale training with vast amount of images.  ...  In sum, our contributions can be summarized as: -We propose a two-sided distance regularization that encourages learning of smooth and mode-preserved latent space through controlled latent interpolation  ... 
arXiv:2111.11672v2 fatcat:eksda5na4jatdlyu47qyjqzqti

Interpolated Joint Space Adversarial Training for Robust and Generalizable Defenses [article]

Chun Pong Lau, Jiang Liu, Hossein Souri, Wei-An Lin, Soheil Feizi, Rama Chellappa
2021 arXiv   pre-print
Recent works show generalization improvement with adversarial samples under novel threat models such as on-manifold threat model or neural perceptual threat model.  ...  To tackle this issue, we propose the Robust Mixup strategy in which we maximize the adversity of the interpolated images and gain robustness and prevent overfitting.  ...  The manifold mixup as a regularization tool uses the interpolation of the hidden representations of a randomly selected layer [50] .  ... 
arXiv:2112.06323v1 fatcat:zclgklsqxrbo7fnuj6c662k4ka

Interpolation Consistency Training for Semi-supervised Learning

Vikas Verma, Alex Lamb, Juho Kannala, Yoshua Bengio, David Lopez-Paz
2019 Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence  
ICT encourages the prediction at an interpolation of unlabeled points to be consistent with the interpolation of the predictions at those points.  ...  We introduce Interpolation Consistency Training (ICT), a simple and computation efficient algorithm for training Deep Neural Networks in the semi-supervised learning paradigm.  ...  In section 2.3, I describe how adversarial training leads to poor performance on unperturbed samples and the justification and summary of Interpolated Adversarial Training (Publication III), which is a  ... 
doi:10.24963/ijcai.2019/504 dblp:conf/ijcai/VermaLKBL19 fatcat:bucjagfaybei5boup7b56yqngu

Suppressing Mislabeled Data via Grouping and Self-Attention [article]

Xiaojiang Peng, Kai Wang, Zhaoyang Zeng, Qing Li, Jianfei Yang, Yu Qiao
2020 arXiv   pre-print
To suppressing the impact of mislabeled data, this paper proposes a conceptually simple yet efficient training block, termed as Attentive Feature Mixup (AFM), which allows paying more attention to clean  ...  Specifically, this plug-and-play AFM first leverages a group-to-attend module to construct groups and assign attention weights for group-wise samples, and then uses a mixup module with the attention weights  ...  In this way, our AFM can be viewed as a regularizer over the training data by massive interpolations.  ... 
arXiv:2010.15603v1 fatcat:zkm7py2bgjg45crdvl5dw7wzdq

NoisyMix: Boosting Model Robustness to Common Corruptions [article]

N. Benjamin Erichson, Soon Hoe Lim, Winnie Xu, Francisco Utrera, Ziang Cao, Michael W. Mahoney
2022 arXiv   pre-print
We demonstrate the benefits of NoisyMix on a range of benchmark datasets, including ImageNet-C, ImageNet-R, and ImageNet-P.  ...  Motivated by this, we introduce NoisyMix, a novel training scheme that promotes stability as well as leverages noisy augmentations in input and feature space to improve both model robustness and in-domain  ...  In contrast, NFM without stability training performs better than the baseline and Manifold Mixup, but not as well as NoisyMix and Manifold Mixup with stability training. (a) All data points.  ... 
arXiv:2202.01263v2 fatcat:qtuh5hv44fan5eywoo46ifttwy

RecursiveMix: Mixed Learning with History [article]

Lingfeng Yang, Xiang Li, Borui Zhao, Renjie Song, Jian Yang
2022 arXiv   pre-print
Based on ResNet-50, RM largely improves classification accuracy by ∼3.2% on CIFAR100 and ∼2.8% on ImageNet with negligible extra computation/storage costs.  ...  In this paper, we propose a recursive mixed-sample learning paradigm, termed "RecursiveMix" (RM), by exploring a novel training strategy that leverages the historical input-prediction-label triplets.  ...  Manifold Mixup [52] , PatchUp [17] and MoEx [36] perform a feature-level interpolation over two samples to prevent overfitting the intermediate representations.  ... 
arXiv:2203.06844v1 fatcat:c5aolvp5vvbplklale7atnpw6i

Learning Temporally Invariant and Localizable Features via Data Augmentation for Video Recognition [article]

Taeoh Kim, Hyeongmin Lee, MyeongAh Cho, Ho Seong Lee, Dong Heon Cho, Sangyoun Lee
2020 arXiv   pre-print
Data augmentation based on visual inductive priors, such as cropping, flipping, rotating, or photometric jittering, is a representative approach to achieve these features.  ...  Recent state-of-the-art recognition solutions have relied on modern data augmentation strategies that exploit a mixture of augmentation operations.  ...  Manifold-MixUp [45] propose a mixing strategy like MixUp but is used instead in the feature space.  ... 
arXiv:2008.05721v1 fatcat:s4crrbntjvbtxjafnt2jj6h63q

Feature Statistics Mixing Regularization for Generative Adversarial Networks [article]

Junho Kim, Yunjey Choi, Youngjung Uh
2022 arXiv   pre-print
As a remedy, we propose feature statistics mixing regularization (FSMR) that encourages the discriminator's prediction to be invariant to the styles of input images.  ...  Specifically, we generate a mixed feature of an original and a reference image in the discriminator's feature space and we apply regularization so that the prediction for the mixed feature is consistent  ...  All experiments were conducted on NAVER Smart Machine Learning (NSML) platform [24, 34] . This work was partly supported by an IITP grant (No.2021-0-00155) and an NRF grant (NRF-2021R1G1A1095637).  ... 
arXiv:2112.04120v2 fatcat:4s6ty4ui4nag7ksybvmcyxq7ri

MUM : Mix Image Tiles and UnMix Feature Tiles for Semi-Supervised Object Detection [article]

JongMok Kim, Jooyoung Jang, Seunghyeon Seo, Jisoo Jeong, Jongkeun Na, Nojun Kwak
2022 arXiv   pre-print
Thus, MUM can enjoy the interpolation-regularization effect from non-interpolated pseudo-labels and successfully generate a meaningful weak-strong pair.  ...  Especially when extending SSL to semi-supervised object detection (SSOD), many strong augmentation methodologies related to image geometry and interpolation-regularization are hard to utilize since they  ...  To summarize, while Cutout [11] has a weak regularization effect, Mixup [46] has the class ambiguity issue in the interpolated label generation process.  ... 
arXiv:2111.10958v2 fatcat:bidangp525aedaxh3sprx4cpgy

An overview of mixing augmentation methods and augmentation strategies [article]

Dominik Lewy, Jacek Mańdziuk
2022 arXiv   pre-print
An alternative approach is application of data augmentation (DA) techniques that aim at model regularization by creating additional observations from the available ones.  ...  Various quantitative comparisons are also included based on the results reported in recent DA literature.  ...  The other method, Co-Mixup, works on the whole mini-batch and attempts to optimize the exposed saliency, as well as encourage diversity among created (augmented) samples.  ... 
arXiv:2107.09887v2 fatcat:isue7dmwxzdihgwiq2efj3k3qm

Adversarial Logit Pairing [article]

Harini Kannan, Alexey Kurakin, Ian Goodfellow
2018 arXiv   pre-print
Next, we introduce enhanced defenses using a technique we call logit pairing, a method that encourages logits for pairs of examples to be similar.  ...  With this new accuracy drop, adversarial logit pairing ties with Tramer et al.(2018) for the state of the art on black box attacks on ImageNet.  ...  Acknowledgements We thank Tom Brown for helpful feedback on drafts of this article.  ... 
arXiv:1803.06373v1 fatcat:7k6sv6623fbnfmyvl33ajhpjqm

SketchTransfer: A Challenging New Task for Exploring Detail-Invariance and the Abstractions Learned by Deep Networks [article]

Alex Lamb, Sherjil Ozair, Vikas Verma, David Ha
2019 arXiv   pre-print
87% accuracy of a classifier trained directly on labeled sketches.  ...  As another example, 3D rendering software is a relatively recent development, yet people are able to understand such rendered scenes even though they are missing details (consider a film like Toy Story  ...  the art regularizer for deep networks which involves training with linear interpolations of images and using a corresponding interpolation of the labels (Equation 1).  ... 
arXiv:1912.11570v1 fatcat:abr4qxuso5fjtmkmcgsjhfhrwq
« Previous Showing results 1 — 15 out of 45 results