SlideShare a Scribd company logo
CutMix
Regularization Strategy to Train Strong Classifiers with Localizable Features
Summary
1. CutMix is extremely simple, requiring
only 20 lines of code.
2. The image on the right shows what
CutMix is and how effective it is.
3. CutMix was performed in conjunction
with other standard data augmentations.
The Code
Introduction
• Data Augmentation is necessary to prevent the Network from focusing on
only a few features and overfitting to them.
• Regional dropout strategies have proven effective on improving
classification and localization tasks by focusing the model on less
discriminative features of images.
• However, they are inefficient as they zero-out or fill regions with random
noise, a serious problem for data-hungry CNNs.
CutMix: Method
• Instead of simply removing pixels replace
the removed regions with a patch from
another image.
• The ground truth labels are also mixed
proportionally to the number of pixels of
combined images.
Advantages
• There is no uninformative pixel during training, making training efficient.
• Attends to non-discriminative parts of objects.
• Enhance localization ability by identifying the object from a partial view.
• Robust to adversarial attacks and Out of Distribution (OOD) examples.
• The training and budgets remains the same. (No extra computation!)
Algorithm
• Let 𝑥 denote a training image and 𝑦 denote its label.
• Generate 𝑥, 𝑦 by combining 𝑥 𝐴, 𝑦 𝐴 and 𝑥 𝐵, 𝑦 𝐵 .
• 𝑥 = 𝑀 ⊙ 𝑥 𝐴 + 1 − 𝑀 ⊙ 𝑥 𝐵
• 𝑦 = 𝜆𝑦 𝐴 + 1 − 𝜆 𝑦 𝐵
• 𝜆 ∼ 𝐵𝑒𝑡𝑎 𝛼, 𝛼 , 𝛼 = 1 for all experiments. 𝜆 is the mixing ratio.
• 𝑀 ∈ {0,1} is a binary mask whose area is determined by 𝜆.
• ⊙ indicates element-wise multiplication.
Mask Generation
• Generate a bounding box 𝐵 = (𝑟𝑥, 𝑟𝑦, 𝑟𝑤, 𝑟ℎ) for the mask.
• 𝑟𝑥~𝑈𝑛𝑖𝑓 0, 𝑊 , 𝑟𝑤 = 𝑊 1 − 𝜆, 𝑊 = 𝑖𝑚𝑎𝑔𝑒 𝑤𝑖𝑑𝑡ℎ
• 𝑟𝑦~𝑈𝑛𝑖𝑓 0, 𝐻 , 𝑟ℎ = 𝐻 1 − 𝜆, 𝐻 = 𝑖𝑚𝑎𝑔𝑒 ℎ𝑒𝑖𝑔ℎ𝑡
• Boundary box is unaware of image size and may not be entirely on the image.
• Fill the mask 𝑀 with 0’s inside 𝐵 and with 1’s outside.
• The cropped area ratio should be 1 − 𝜆.
• Rounding error and boundary box overflow may cause the true 𝜆 to be different.
Results
Comparison with Other Methods
One can see that CutMix learns to identify separate parts of
the image by analyzing the CAM visualizations.
Other methods either cannot do this (Mixup) or are
inefficient (Cutout).
Performance on Classification
Resistance to Overfitting
Weakly Supervised Object Localization
• Weakly Supervised Object
Localization (WSOL) is a task
where the aim is to localize the
object when only the class label
is available.
Transfer Learning of CutMix Models
Ablation Studies
• Ablation study using different values of
𝛼 on the beta distribution shows that 𝛼 =
1 has the best results.
• Ablation study with using CutMix at
different layers shows that using CutMix
on the input layer produces the best
results.
• Different implementations of CutMix are
all inferior to the proposed
implementation.
Robustness
• Deep Learning models are prone to overconfident predictions.
• This leads them to be vulnerable to adversarial attacks.
• They are also prone to giving incorrect answers when the test data is
from a different distribution from the training data.
Robustness to Image Manipulation
Robustness to Different Images
Conclusion
• CutMix is a simple yet effective regularization technique that
improves performance on a wide array of tasks.
• By forcing the model to focus on non-discriminative parts of an image
efficiently, it both improves performance and reduces overconfident
predictions.
Final Thoughts
• Perhaps using CutMix in GANs will reduce mode collapse and
improve performance.
• Image Segmentation may also benefit from CutMix.

More Related Content

What's hot

Image segmentation 2
Image segmentation 2 Image segmentation 2
Image segmentation 2 Rumah Belajar
 
Image segmentation
Image segmentationImage segmentation
Image segmentation
khyati gupta
 
Regularization
RegularizationRegularization
Regularization
Darren Yow-Bang Wang
 
Feature Extraction
Feature ExtractionFeature Extraction
Feature Extractionskylian
 
Real Time Object Dectection using machine learning
Real Time Object Dectection using machine learningReal Time Object Dectection using machine learning
Real Time Object Dectection using machine learning
pratik pratyay
 
[DL輪読会]High-Fidelity Image Generation with Fewer Labels
[DL輪読会]High-Fidelity Image Generation with Fewer Labels[DL輪読会]High-Fidelity Image Generation with Fewer Labels
[DL輪読会]High-Fidelity Image Generation with Fewer Labels
Deep Learning JP
 
はじめてのパターン認識 第11章 11.1-11.2
はじめてのパターン認識 第11章 11.1-11.2はじめてのパターン認識 第11章 11.1-11.2
はじめてのパターン認識 第11章 11.1-11.2
Prunus 1350
 
[DL輪読会]NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Det...
[DL輪読会]NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Det...[DL輪読会]NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Det...
[DL輪読会]NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Det...
Deep Learning JP
 
Feature Pyramid Network, FPN
Feature Pyramid Network, FPNFeature Pyramid Network, FPN
Feature Pyramid Network, FPN
Institute of Agricultural Machinery, NARO
 
Fisher線形判別分析とFisher Weight Maps
Fisher線形判別分析とFisher Weight MapsFisher線形判別分析とFisher Weight Maps
Fisher線形判別分析とFisher Weight Maps
Takao Yamanaka
 
Anchor free object detection by deep learning
Anchor free object detection by deep learningAnchor free object detection by deep learning
Anchor free object detection by deep learning
Yu Huang
 
Decision tree
Decision treeDecision tree
Decision tree
SEMINARGROOT
 
A study and comparison of different image segmentation algorithms
A study and comparison of different image segmentation algorithmsA study and comparison of different image segmentation algorithms
A study and comparison of different image segmentation algorithms
Manje Gowda
 
Object detection - RCNNs vs Retinanet
Object detection - RCNNs vs RetinanetObject detection - RCNNs vs Retinanet
Object detection - RCNNs vs Retinanet
Rishabh Indoria
 
The world of loss function
The world of loss functionThe world of loss function
The world of loss function
홍배 김
 
Support Vector Machines (SVM)
Support Vector Machines (SVM)Support Vector Machines (SVM)
Support Vector Machines (SVM)
FAO
 
(2022年3月版)深層学習によるImage Classificaitonの発展
(2022年3月版)深層学習によるImage Classificaitonの発展(2022年3月版)深層学習によるImage Classificaitonの発展
(2022年3月版)深層学習によるImage Classificaitonの発展
Takumi Ohkuma
 
Neural Networks
Neural NetworksNeural Networks
Neural Networks
Ismail El Gayar
 
Image segmentation
Image segmentationImage segmentation
Image segmentationDeepak Kumar
 

What's hot (20)

Image segmentation 2
Image segmentation 2 Image segmentation 2
Image segmentation 2
 
Image segmentation
Image segmentationImage segmentation
Image segmentation
 
Regularization
RegularizationRegularization
Regularization
 
Feature Extraction
Feature ExtractionFeature Extraction
Feature Extraction
 
Real Time Object Dectection using machine learning
Real Time Object Dectection using machine learningReal Time Object Dectection using machine learning
Real Time Object Dectection using machine learning
 
[DL輪読会]High-Fidelity Image Generation with Fewer Labels
[DL輪読会]High-Fidelity Image Generation with Fewer Labels[DL輪読会]High-Fidelity Image Generation with Fewer Labels
[DL輪読会]High-Fidelity Image Generation with Fewer Labels
 
はじめてのパターン認識 第11章 11.1-11.2
はじめてのパターン認識 第11章 11.1-11.2はじめてのパターン認識 第11章 11.1-11.2
はじめてのパターン認識 第11章 11.1-11.2
 
[DL輪読会]NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Det...
[DL輪読会]NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Det...[DL輪読会]NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Det...
[DL輪読会]NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Det...
 
Feature Pyramid Network, FPN
Feature Pyramid Network, FPNFeature Pyramid Network, FPN
Feature Pyramid Network, FPN
 
Fisher線形判別分析とFisher Weight Maps
Fisher線形判別分析とFisher Weight MapsFisher線形判別分析とFisher Weight Maps
Fisher線形判別分析とFisher Weight Maps
 
Anchor free object detection by deep learning
Anchor free object detection by deep learningAnchor free object detection by deep learning
Anchor free object detection by deep learning
 
Decision tree
Decision treeDecision tree
Decision tree
 
A study and comparison of different image segmentation algorithms
A study and comparison of different image segmentation algorithmsA study and comparison of different image segmentation algorithms
A study and comparison of different image segmentation algorithms
 
Object detection - RCNNs vs Retinanet
Object detection - RCNNs vs RetinanetObject detection - RCNNs vs Retinanet
Object detection - RCNNs vs Retinanet
 
The world of loss function
The world of loss functionThe world of loss function
The world of loss function
 
Support Vector Machines (SVM)
Support Vector Machines (SVM)Support Vector Machines (SVM)
Support Vector Machines (SVM)
 
(2022年3月版)深層学習によるImage Classificaitonの発展
(2022年3月版)深層学習によるImage Classificaitonの発展(2022年3月版)深層学習によるImage Classificaitonの発展
(2022年3月版)深層学習によるImage Classificaitonの発展
 
Neural Networks
Neural NetworksNeural Networks
Neural Networks
 
Image segmentation
Image segmentationImage segmentation
Image segmentation
 
Lecture11 - neural networks
Lecture11 - neural networksLecture11 - neural networks
Lecture11 - neural networks
 

Similar to CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features

Hands-On Machine Learning with Scikit-Learn and TensorFlow - Chapter8
Hands-On Machine Learning with Scikit-Learn and TensorFlow - Chapter8Hands-On Machine Learning with Scikit-Learn and TensorFlow - Chapter8
Hands-On Machine Learning with Scikit-Learn and TensorFlow - Chapter8
Hakky St
 
PR095: Modularity Matters: Learning Invariant Relational Reasoning Tasks
PR095: Modularity Matters: Learning Invariant Relational Reasoning TasksPR095: Modularity Matters: Learning Invariant Relational Reasoning Tasks
PR095: Modularity Matters: Learning Invariant Relational Reasoning Tasks
Jinwon Lee
 
Mnist soln
Mnist solnMnist soln
Mnist soln
DanishFaisal4
 
Overview of Convolutional Neural Networks
Overview of Convolutional Neural NetworksOverview of Convolutional Neural Networks
Overview of Convolutional Neural Networks
ananth
 
MLSEV Virtual. State of the Art in ML
MLSEV Virtual. State of the Art in MLMLSEV Virtual. State of the Art in ML
MLSEV Virtual. State of the Art in ML
BigML, Inc
 
J. Park, AAAI 2022, MLILAB, KAIST AI
J. Park, AAAI 2022, MLILAB, KAIST AIJ. Park, AAAI 2022, MLILAB, KAIST AI
J. Park, AAAI 2022, MLILAB, KAIST AI
MLILAB
 
BMVA summer school MATLAB programming tutorial
BMVA summer school MATLAB programming tutorialBMVA summer school MATLAB programming tutorial
BMVA summer school MATLAB programming tutorial
potaters
 
PR-231: A Simple Framework for Contrastive Learning of Visual Representations
PR-231: A Simple Framework for Contrastive Learning of Visual RepresentationsPR-231: A Simple Framework for Contrastive Learning of Visual Representations
PR-231: A Simple Framework for Contrastive Learning of Visual Representations
Jinwon Lee
 
Rethinking Data Augmentation for Image Super-resolution: A Comprehensive Anal...
Rethinking Data Augmentation for Image Super-resolution: A Comprehensive Anal...Rethinking Data Augmentation for Image Super-resolution: A Comprehensive Anal...
Rethinking Data Augmentation for Image Super-resolution: A Comprehensive Anal...
JaeJun Yoo
 
SIRG-BSU_3_used-important.pdf
SIRG-BSU_3_used-important.pdfSIRG-BSU_3_used-important.pdf
SIRG-BSU_3_used-important.pdf
DrAhmedElngar
 
Kulakov_Mykhailo.pptx
Kulakov_Mykhailo.pptxKulakov_Mykhailo.pptx
Kulakov_Mykhailo.pptx
EE22S501AlavalaNagaS
 
Handwritten Digit Recognition and performance of various modelsation[autosaved]
Handwritten Digit Recognition and performance of various modelsation[autosaved]Handwritten Digit Recognition and performance of various modelsation[autosaved]
Handwritten Digit Recognition and performance of various modelsation[autosaved]
SubhradeepMaji
 
Face Detection.pptx
Face Detection.pptxFace Detection.pptx
Face Detection.pptx
TorshaSett
 
Digital Image Processing.pptx
Digital Image Processing.pptxDigital Image Processing.pptx
Digital Image Processing.pptx
Aishwarya SenthilNathan
 
Deep learning
Deep learningDeep learning
Deep learning
Rouyun Pan
 
Jaemin_230701_Simple_Copy_paste.pptx
Jaemin_230701_Simple_Copy_paste.pptxJaemin_230701_Simple_Copy_paste.pptx
Jaemin_230701_Simple_Copy_paste.pptx
JAEMINJEONG5
 
lec6a.ppt
lec6a.pptlec6a.ppt
lec6a.ppt
SaadMemon23
 
Exploring Simple Siamese Representation Learning
Exploring Simple Siamese Representation LearningExploring Simple Siamese Representation Learning
Exploring Simple Siamese Representation Learning
Sangmin Woo
 
Deep learning and its application
Deep learning and its applicationDeep learning and its application
Deep learning and its application
Srishty Saha
 
Seeing is Not Believing: Camouflage Attacks on Image Scaling Algorithms Review
Seeing is Not Believing: Camouflage Attacks on Image Scaling Algorithms ReviewSeeing is Not Believing: Camouflage Attacks on Image Scaling Algorithms Review
Seeing is Not Believing: Camouflage Attacks on Image Scaling Algorithms Review
Junyaup Kim
 

Similar to CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features (20)

Hands-On Machine Learning with Scikit-Learn and TensorFlow - Chapter8
Hands-On Machine Learning with Scikit-Learn and TensorFlow - Chapter8Hands-On Machine Learning with Scikit-Learn and TensorFlow - Chapter8
Hands-On Machine Learning with Scikit-Learn and TensorFlow - Chapter8
 
PR095: Modularity Matters: Learning Invariant Relational Reasoning Tasks
PR095: Modularity Matters: Learning Invariant Relational Reasoning TasksPR095: Modularity Matters: Learning Invariant Relational Reasoning Tasks
PR095: Modularity Matters: Learning Invariant Relational Reasoning Tasks
 
Mnist soln
Mnist solnMnist soln
Mnist soln
 
Overview of Convolutional Neural Networks
Overview of Convolutional Neural NetworksOverview of Convolutional Neural Networks
Overview of Convolutional Neural Networks
 
MLSEV Virtual. State of the Art in ML
MLSEV Virtual. State of the Art in MLMLSEV Virtual. State of the Art in ML
MLSEV Virtual. State of the Art in ML
 
J. Park, AAAI 2022, MLILAB, KAIST AI
J. Park, AAAI 2022, MLILAB, KAIST AIJ. Park, AAAI 2022, MLILAB, KAIST AI
J. Park, AAAI 2022, MLILAB, KAIST AI
 
BMVA summer school MATLAB programming tutorial
BMVA summer school MATLAB programming tutorialBMVA summer school MATLAB programming tutorial
BMVA summer school MATLAB programming tutorial
 
PR-231: A Simple Framework for Contrastive Learning of Visual Representations
PR-231: A Simple Framework for Contrastive Learning of Visual RepresentationsPR-231: A Simple Framework for Contrastive Learning of Visual Representations
PR-231: A Simple Framework for Contrastive Learning of Visual Representations
 
Rethinking Data Augmentation for Image Super-resolution: A Comprehensive Anal...
Rethinking Data Augmentation for Image Super-resolution: A Comprehensive Anal...Rethinking Data Augmentation for Image Super-resolution: A Comprehensive Anal...
Rethinking Data Augmentation for Image Super-resolution: A Comprehensive Anal...
 
SIRG-BSU_3_used-important.pdf
SIRG-BSU_3_used-important.pdfSIRG-BSU_3_used-important.pdf
SIRG-BSU_3_used-important.pdf
 
Kulakov_Mykhailo.pptx
Kulakov_Mykhailo.pptxKulakov_Mykhailo.pptx
Kulakov_Mykhailo.pptx
 
Handwritten Digit Recognition and performance of various modelsation[autosaved]
Handwritten Digit Recognition and performance of various modelsation[autosaved]Handwritten Digit Recognition and performance of various modelsation[autosaved]
Handwritten Digit Recognition and performance of various modelsation[autosaved]
 
Face Detection.pptx
Face Detection.pptxFace Detection.pptx
Face Detection.pptx
 
Digital Image Processing.pptx
Digital Image Processing.pptxDigital Image Processing.pptx
Digital Image Processing.pptx
 
Deep learning
Deep learningDeep learning
Deep learning
 
Jaemin_230701_Simple_Copy_paste.pptx
Jaemin_230701_Simple_Copy_paste.pptxJaemin_230701_Simple_Copy_paste.pptx
Jaemin_230701_Simple_Copy_paste.pptx
 
lec6a.ppt
lec6a.pptlec6a.ppt
lec6a.ppt
 
Exploring Simple Siamese Representation Learning
Exploring Simple Siamese Representation LearningExploring Simple Siamese Representation Learning
Exploring Simple Siamese Representation Learning
 
Deep learning and its application
Deep learning and its applicationDeep learning and its application
Deep learning and its application
 
Seeing is Not Believing: Camouflage Attacks on Image Scaling Algorithms Review
Seeing is Not Believing: Camouflage Attacks on Image Scaling Algorithms ReviewSeeing is Not Believing: Camouflage Attacks on Image Scaling Algorithms Review
Seeing is Not Believing: Camouflage Attacks on Image Scaling Algorithms Review
 

More from Joonhyung Lee

nnUNet
nnUNetnnUNet
Rethinking Attention with Performers
Rethinking Attention with PerformersRethinking Attention with Performers
Rethinking Attention with Performers
Joonhyung Lee
 
Denoising Unpaired Low Dose CT Images with Self-Ensembled CycleGAN
Denoising Unpaired Low Dose CT Images with Self-Ensembled CycleGANDenoising Unpaired Low Dose CT Images with Self-Ensembled CycleGAN
Denoising Unpaired Low Dose CT Images with Self-Ensembled CycleGAN
Joonhyung Lee
 
Deep Learning Fast MRI Using Channel Attention in Magnitude Domain
Deep Learning Fast MRI Using Channel Attention in Magnitude DomainDeep Learning Fast MRI Using Channel Attention in Magnitude Domain
Deep Learning Fast MRI Using Channel Attention in Magnitude Domain
Joonhyung Lee
 
InfoGAN: Interpretable Representation Learning by Information Maximizing Gene...
InfoGAN: Interpretable Representation Learning by Information Maximizing Gene...InfoGAN: Interpretable Representation Learning by Information Maximizing Gene...
InfoGAN: Interpretable Representation Learning by Information Maximizing Gene...
Joonhyung Lee
 
AlphaZero: A General Reinforcement Learning Algorithm that Masters Chess, Sho...
AlphaZero: A General Reinforcement Learning Algorithm that Masters Chess, Sho...AlphaZero: A General Reinforcement Learning Algorithm that Masters Chess, Sho...
AlphaZero: A General Reinforcement Learning Algorithm that Masters Chess, Sho...
Joonhyung Lee
 
Squeeze Excitation Networks, The simple idea that won the final ImageNet Chal...
Squeeze Excitation Networks, The simple idea that won the final ImageNet Chal...Squeeze Excitation Networks, The simple idea that won the final ImageNet Chal...
Squeeze Excitation Networks, The simple idea that won the final ImageNet Chal...
Joonhyung Lee
 
DeepLab V3+: Encoder-Decoder with Atrous Separable Convolution for Semantic I...
DeepLab V3+: Encoder-Decoder with Atrous Separable Convolution for Semantic I...DeepLab V3+: Encoder-Decoder with Atrous Separable Convolution for Semantic I...
DeepLab V3+: Encoder-Decoder with Atrous Separable Convolution for Semantic I...
Joonhyung Lee
 
AlphaGo Zero: Mastering the Game of Go Without Human Knowledge
AlphaGo Zero: Mastering the Game of Go Without Human KnowledgeAlphaGo Zero: Mastering the Game of Go Without Human Knowledge
AlphaGo Zero: Mastering the Game of Go Without Human Knowledge
Joonhyung Lee
 
StarGAN
StarGANStarGAN
StarGAN
Joonhyung Lee
 
Deep Learning in Bio-Medical Imaging
Deep Learning in Bio-Medical ImagingDeep Learning in Bio-Medical Imaging
Deep Learning in Bio-Medical Imaging
Joonhyung Lee
 

More from Joonhyung Lee (11)

nnUNet
nnUNetnnUNet
nnUNet
 
Rethinking Attention with Performers
Rethinking Attention with PerformersRethinking Attention with Performers
Rethinking Attention with Performers
 
Denoising Unpaired Low Dose CT Images with Self-Ensembled CycleGAN
Denoising Unpaired Low Dose CT Images with Self-Ensembled CycleGANDenoising Unpaired Low Dose CT Images with Self-Ensembled CycleGAN
Denoising Unpaired Low Dose CT Images with Self-Ensembled CycleGAN
 
Deep Learning Fast MRI Using Channel Attention in Magnitude Domain
Deep Learning Fast MRI Using Channel Attention in Magnitude DomainDeep Learning Fast MRI Using Channel Attention in Magnitude Domain
Deep Learning Fast MRI Using Channel Attention in Magnitude Domain
 
InfoGAN: Interpretable Representation Learning by Information Maximizing Gene...
InfoGAN: Interpretable Representation Learning by Information Maximizing Gene...InfoGAN: Interpretable Representation Learning by Information Maximizing Gene...
InfoGAN: Interpretable Representation Learning by Information Maximizing Gene...
 
AlphaZero: A General Reinforcement Learning Algorithm that Masters Chess, Sho...
AlphaZero: A General Reinforcement Learning Algorithm that Masters Chess, Sho...AlphaZero: A General Reinforcement Learning Algorithm that Masters Chess, Sho...
AlphaZero: A General Reinforcement Learning Algorithm that Masters Chess, Sho...
 
Squeeze Excitation Networks, The simple idea that won the final ImageNet Chal...
Squeeze Excitation Networks, The simple idea that won the final ImageNet Chal...Squeeze Excitation Networks, The simple idea that won the final ImageNet Chal...
Squeeze Excitation Networks, The simple idea that won the final ImageNet Chal...
 
DeepLab V3+: Encoder-Decoder with Atrous Separable Convolution for Semantic I...
DeepLab V3+: Encoder-Decoder with Atrous Separable Convolution for Semantic I...DeepLab V3+: Encoder-Decoder with Atrous Separable Convolution for Semantic I...
DeepLab V3+: Encoder-Decoder with Atrous Separable Convolution for Semantic I...
 
AlphaGo Zero: Mastering the Game of Go Without Human Knowledge
AlphaGo Zero: Mastering the Game of Go Without Human KnowledgeAlphaGo Zero: Mastering the Game of Go Without Human Knowledge
AlphaGo Zero: Mastering the Game of Go Without Human Knowledge
 
StarGAN
StarGANStarGAN
StarGAN
 
Deep Learning in Bio-Medical Imaging
Deep Learning in Bio-Medical ImagingDeep Learning in Bio-Medical Imaging
Deep Learning in Bio-Medical Imaging
 

Recently uploaded

The Future of Platform Engineering
The Future of Platform EngineeringThe Future of Platform Engineering
The Future of Platform Engineering
Jemma Hussein Allen
 
From Daily Decisions to Bottom Line: Connecting Product Work to Revenue by VP...
From Daily Decisions to Bottom Line: Connecting Product Work to Revenue by VP...From Daily Decisions to Bottom Line: Connecting Product Work to Revenue by VP...
From Daily Decisions to Bottom Line: Connecting Product Work to Revenue by VP...
Product School
 
GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using Deplo...
GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using Deplo...GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using Deplo...
GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using Deplo...
James Anderson
 
Assuring Contact Center Experiences for Your Customers With ThousandEyes
Assuring Contact Center Experiences for Your Customers With ThousandEyesAssuring Contact Center Experiences for Your Customers With ThousandEyes
Assuring Contact Center Experiences for Your Customers With ThousandEyes
ThousandEyes
 
FIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdf
FIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdfFIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdf
FIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdf
FIDO Alliance
 
Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...
Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...
Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...
Thierry Lestable
 
The Art of the Pitch: WordPress Relationships and Sales
The Art of the Pitch: WordPress Relationships and SalesThe Art of the Pitch: WordPress Relationships and Sales
The Art of the Pitch: WordPress Relationships and Sales
Laura Byrne
 
How world-class product teams are winning in the AI era by CEO and Founder, P...
How world-class product teams are winning in the AI era by CEO and Founder, P...How world-class product teams are winning in the AI era by CEO and Founder, P...
How world-class product teams are winning in the AI era by CEO and Founder, P...
Product School
 
Key Trends Shaping the Future of Infrastructure.pdf
Key Trends Shaping the Future of Infrastructure.pdfKey Trends Shaping the Future of Infrastructure.pdf
Key Trends Shaping the Future of Infrastructure.pdf
Cheryl Hung
 
AI for Every Business: Unlocking Your Product's Universal Potential by VP of ...
AI for Every Business: Unlocking Your Product's Universal Potential by VP of ...AI for Every Business: Unlocking Your Product's Universal Potential by VP of ...
AI for Every Business: Unlocking Your Product's Universal Potential by VP of ...
Product School
 
De-mystifying Zero to One: Design Informed Techniques for Greenfield Innovati...
De-mystifying Zero to One: Design Informed Techniques for Greenfield Innovati...De-mystifying Zero to One: Design Informed Techniques for Greenfield Innovati...
De-mystifying Zero to One: Design Informed Techniques for Greenfield Innovati...
Product School
 
JMeter webinar - integration with InfluxDB and Grafana
JMeter webinar - integration with InfluxDB and GrafanaJMeter webinar - integration with InfluxDB and Grafana
JMeter webinar - integration with InfluxDB and Grafana
RTTS
 
Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024
Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024
Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024
Tobias Schneck
 
Essentials of Automations: Optimizing FME Workflows with Parameters
Essentials of Automations: Optimizing FME Workflows with ParametersEssentials of Automations: Optimizing FME Workflows with Parameters
Essentials of Automations: Optimizing FME Workflows with Parameters
Safe Software
 
"Impact of front-end architecture on development cost", Viktor Turskyi
"Impact of front-end architecture on development cost", Viktor Turskyi"Impact of front-end architecture on development cost", Viktor Turskyi
"Impact of front-end architecture on development cost", Viktor Turskyi
Fwdays
 
GenAISummit 2024 May 28 Sri Ambati Keynote: AGI Belongs to The Community in O...
GenAISummit 2024 May 28 Sri Ambati Keynote: AGI Belongs to The Community in O...GenAISummit 2024 May 28 Sri Ambati Keynote: AGI Belongs to The Community in O...
GenAISummit 2024 May 28 Sri Ambati Keynote: AGI Belongs to The Community in O...
Sri Ambati
 
Unsubscribed: Combat Subscription Fatigue With a Membership Mentality by Head...
Unsubscribed: Combat Subscription Fatigue With a Membership Mentality by Head...Unsubscribed: Combat Subscription Fatigue With a Membership Mentality by Head...
Unsubscribed: Combat Subscription Fatigue With a Membership Mentality by Head...
Product School
 
FIDO Alliance Osaka Seminar: Overview.pdf
FIDO Alliance Osaka Seminar: Overview.pdfFIDO Alliance Osaka Seminar: Overview.pdf
FIDO Alliance Osaka Seminar: Overview.pdf
FIDO Alliance
 
DevOps and Testing slides at DASA Connect
DevOps and Testing slides at DASA ConnectDevOps and Testing slides at DASA Connect
DevOps and Testing slides at DASA Connect
Kari Kakkonen
 
FIDO Alliance Osaka Seminar: FIDO Security Aspects.pdf
FIDO Alliance Osaka Seminar: FIDO Security Aspects.pdfFIDO Alliance Osaka Seminar: FIDO Security Aspects.pdf
FIDO Alliance Osaka Seminar: FIDO Security Aspects.pdf
FIDO Alliance
 

Recently uploaded (20)

The Future of Platform Engineering
The Future of Platform EngineeringThe Future of Platform Engineering
The Future of Platform Engineering
 
From Daily Decisions to Bottom Line: Connecting Product Work to Revenue by VP...
From Daily Decisions to Bottom Line: Connecting Product Work to Revenue by VP...From Daily Decisions to Bottom Line: Connecting Product Work to Revenue by VP...
From Daily Decisions to Bottom Line: Connecting Product Work to Revenue by VP...
 
GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using Deplo...
GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using Deplo...GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using Deplo...
GDG Cloud Southlake #33: Boule & Rebala: Effective AppSec in SDLC using Deplo...
 
Assuring Contact Center Experiences for Your Customers With ThousandEyes
Assuring Contact Center Experiences for Your Customers With ThousandEyesAssuring Contact Center Experiences for Your Customers With ThousandEyes
Assuring Contact Center Experiences for Your Customers With ThousandEyes
 
FIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdf
FIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdfFIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdf
FIDO Alliance Osaka Seminar: The WebAuthn API and Discoverable Credentials.pdf
 
Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...
Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...
Empowering NextGen Mobility via Large Action Model Infrastructure (LAMI): pav...
 
The Art of the Pitch: WordPress Relationships and Sales
The Art of the Pitch: WordPress Relationships and SalesThe Art of the Pitch: WordPress Relationships and Sales
The Art of the Pitch: WordPress Relationships and Sales
 
How world-class product teams are winning in the AI era by CEO and Founder, P...
How world-class product teams are winning in the AI era by CEO and Founder, P...How world-class product teams are winning in the AI era by CEO and Founder, P...
How world-class product teams are winning in the AI era by CEO and Founder, P...
 
Key Trends Shaping the Future of Infrastructure.pdf
Key Trends Shaping the Future of Infrastructure.pdfKey Trends Shaping the Future of Infrastructure.pdf
Key Trends Shaping the Future of Infrastructure.pdf
 
AI for Every Business: Unlocking Your Product's Universal Potential by VP of ...
AI for Every Business: Unlocking Your Product's Universal Potential by VP of ...AI for Every Business: Unlocking Your Product's Universal Potential by VP of ...
AI for Every Business: Unlocking Your Product's Universal Potential by VP of ...
 
De-mystifying Zero to One: Design Informed Techniques for Greenfield Innovati...
De-mystifying Zero to One: Design Informed Techniques for Greenfield Innovati...De-mystifying Zero to One: Design Informed Techniques for Greenfield Innovati...
De-mystifying Zero to One: Design Informed Techniques for Greenfield Innovati...
 
JMeter webinar - integration with InfluxDB and Grafana
JMeter webinar - integration with InfluxDB and GrafanaJMeter webinar - integration with InfluxDB and Grafana
JMeter webinar - integration with InfluxDB and Grafana
 
Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024
Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024
Kubernetes & AI - Beauty and the Beast !?! @KCD Istanbul 2024
 
Essentials of Automations: Optimizing FME Workflows with Parameters
Essentials of Automations: Optimizing FME Workflows with ParametersEssentials of Automations: Optimizing FME Workflows with Parameters
Essentials of Automations: Optimizing FME Workflows with Parameters
 
"Impact of front-end architecture on development cost", Viktor Turskyi
"Impact of front-end architecture on development cost", Viktor Turskyi"Impact of front-end architecture on development cost", Viktor Turskyi
"Impact of front-end architecture on development cost", Viktor Turskyi
 
GenAISummit 2024 May 28 Sri Ambati Keynote: AGI Belongs to The Community in O...
GenAISummit 2024 May 28 Sri Ambati Keynote: AGI Belongs to The Community in O...GenAISummit 2024 May 28 Sri Ambati Keynote: AGI Belongs to The Community in O...
GenAISummit 2024 May 28 Sri Ambati Keynote: AGI Belongs to The Community in O...
 
Unsubscribed: Combat Subscription Fatigue With a Membership Mentality by Head...
Unsubscribed: Combat Subscription Fatigue With a Membership Mentality by Head...Unsubscribed: Combat Subscription Fatigue With a Membership Mentality by Head...
Unsubscribed: Combat Subscription Fatigue With a Membership Mentality by Head...
 
FIDO Alliance Osaka Seminar: Overview.pdf
FIDO Alliance Osaka Seminar: Overview.pdfFIDO Alliance Osaka Seminar: Overview.pdf
FIDO Alliance Osaka Seminar: Overview.pdf
 
DevOps and Testing slides at DASA Connect
DevOps and Testing slides at DASA ConnectDevOps and Testing slides at DASA Connect
DevOps and Testing slides at DASA Connect
 
FIDO Alliance Osaka Seminar: FIDO Security Aspects.pdf
FIDO Alliance Osaka Seminar: FIDO Security Aspects.pdfFIDO Alliance Osaka Seminar: FIDO Security Aspects.pdf
FIDO Alliance Osaka Seminar: FIDO Security Aspects.pdf
 

CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features

  • 1. CutMix Regularization Strategy to Train Strong Classifiers with Localizable Features
  • 2. Summary 1. CutMix is extremely simple, requiring only 20 lines of code. 2. The image on the right shows what CutMix is and how effective it is. 3. CutMix was performed in conjunction with other standard data augmentations.
  • 4. Introduction • Data Augmentation is necessary to prevent the Network from focusing on only a few features and overfitting to them. • Regional dropout strategies have proven effective on improving classification and localization tasks by focusing the model on less discriminative features of images. • However, they are inefficient as they zero-out or fill regions with random noise, a serious problem for data-hungry CNNs.
  • 5. CutMix: Method • Instead of simply removing pixels replace the removed regions with a patch from another image. • The ground truth labels are also mixed proportionally to the number of pixels of combined images.
  • 6. Advantages • There is no uninformative pixel during training, making training efficient. • Attends to non-discriminative parts of objects. • Enhance localization ability by identifying the object from a partial view. • Robust to adversarial attacks and Out of Distribution (OOD) examples. • The training and budgets remains the same. (No extra computation!)
  • 7. Algorithm • Let 𝑥 denote a training image and 𝑦 denote its label. • Generate 𝑥, 𝑦 by combining 𝑥 𝐴, 𝑦 𝐴 and 𝑥 𝐵, 𝑦 𝐵 . • 𝑥 = 𝑀 ⊙ 𝑥 𝐴 + 1 − 𝑀 ⊙ 𝑥 𝐵 • 𝑦 = 𝜆𝑦 𝐴 + 1 − 𝜆 𝑦 𝐵 • 𝜆 ∼ 𝐵𝑒𝑡𝑎 𝛼, 𝛼 , 𝛼 = 1 for all experiments. 𝜆 is the mixing ratio. • 𝑀 ∈ {0,1} is a binary mask whose area is determined by 𝜆. • ⊙ indicates element-wise multiplication.
  • 8. Mask Generation • Generate a bounding box 𝐵 = (𝑟𝑥, 𝑟𝑦, 𝑟𝑤, 𝑟ℎ) for the mask. • 𝑟𝑥~𝑈𝑛𝑖𝑓 0, 𝑊 , 𝑟𝑤 = 𝑊 1 − 𝜆, 𝑊 = 𝑖𝑚𝑎𝑔𝑒 𝑤𝑖𝑑𝑡ℎ • 𝑟𝑦~𝑈𝑛𝑖𝑓 0, 𝐻 , 𝑟ℎ = 𝐻 1 − 𝜆, 𝐻 = 𝑖𝑚𝑎𝑔𝑒 ℎ𝑒𝑖𝑔ℎ𝑡 • Boundary box is unaware of image size and may not be entirely on the image. • Fill the mask 𝑀 with 0’s inside 𝐵 and with 1’s outside. • The cropped area ratio should be 1 − 𝜆. • Rounding error and boundary box overflow may cause the true 𝜆 to be different.
  • 10. Comparison with Other Methods One can see that CutMix learns to identify separate parts of the image by analyzing the CAM visualizations. Other methods either cannot do this (Mixup) or are inefficient (Cutout).
  • 13. Weakly Supervised Object Localization • Weakly Supervised Object Localization (WSOL) is a task where the aim is to localize the object when only the class label is available.
  • 14. Transfer Learning of CutMix Models
  • 15. Ablation Studies • Ablation study using different values of 𝛼 on the beta distribution shows that 𝛼 = 1 has the best results. • Ablation study with using CutMix at different layers shows that using CutMix on the input layer produces the best results. • Different implementations of CutMix are all inferior to the proposed implementation.
  • 16. Robustness • Deep Learning models are prone to overconfident predictions. • This leads them to be vulnerable to adversarial attacks. • They are also prone to giving incorrect answers when the test data is from a different distribution from the training data.
  • 17. Robustness to Image Manipulation
  • 19. Conclusion • CutMix is a simple yet effective regularization technique that improves performance on a wide array of tasks. • By forcing the model to focus on non-discriminative parts of an image efficiently, it both improves performance and reduces overconfident predictions.
  • 20. Final Thoughts • Perhaps using CutMix in GANs will reduce mode collapse and improve performance. • Image Segmentation may also benefit from CutMix.