DOMAIN-FREE ADVERSARIAL

SPLITTING

FOR

DOMAIN

GENERALIZATION

Presenter : Sai Aravind Sreeramadas

• Introduction
• Vocabulary
• Objective
• Related work
• DAFS
• Main Idea
• Formulation
• Optimisation
• L2 -norm
• Experiments
• MSDS
• SSDS
• SSTD
• Ablation study
• Additional details
• Takeaways
• Open questions and Discussion

​

Overview

Introduction

Domain Generalisation

• Use several source domains to train the learner to generalize to unseen target domain ( To tackle domain shift issue )

Domain Adaptation

• A transferable learner from source domain to target domain.

DG Objective

• To learn the invariant features across domains
• Generalize well to unseen target domain
• To tackle domain shift issue.

​

​

Related work

DG with domain labels.

When the domain labels are available

1. Learn domain invariant features by aligning feature distributions
2. Based on meta-learning approach
3. Augment fake domain data to train learner for enhancing generalization capability of learner.

​

DG without domain labels.

​

When the domain label is unavailable,

​

1. Extracts robust feature representation by projecting out superficial patterns like color and texture.
2. Solve jigsaw puzzles of the training images.
3. Clustering and learn domain invariant features via adversarial training of feature extractor and domain discriminator.
4. Adversarial data augmentation to tackle the setting that the training set comprises a single domain.

DOMAIN-FREE ADVERSARIAL SPLITTING MODEL

What is this trying to do :

This paper aims to learn to generalize well to unseen target domain without relying on the knowledge of the number of source domains and domain labels.

They model DG as a learning problem that enforces the learner to be able to generalize well for any train/val subsets splitting of the training dataset using meta-learning approach

Main Idea

In this domain-free setting,

Setup:

1. Do not assume or know the domains in the training dataset
2. Split train/val subsets, i.e., synthetic source/target domains, over the training dataset
3. Train meta learner over this split.

Hypothesis:

1. Training on those conditions explicitly enforces that the trained learner should be generalizable for any possible domain shifts within the training dataset.
2. It would be generalize even better if this domain shift in train/val is maximised.(Adversarial splitting)

​

Formulation

• Γξ = {Sv : Sv ⊂ S, |Sv| = ξ} is the set of all possible val-subsets of S with length of ξ,
• St = S − Sv is train-subset paired with each Sv,
• L(θ(w); Sv) is the loss on Sv, where θ(w) is the parameters of f,
• L(θ; St, w) is L(θ; St) with θ initialized by w
• R(w) is regularization term

This is the main objective is a min max problem that is being solved to prove their hypothesis for this DG setting.

Optimisation

Since Eq. (2) is a min-max optimization problem, we alternately update Sv and w by fixing the other one as known

Inner loop for optimization of θ(w)

Finding the hardest splitting Sv with fixed w

Fixing w, to find Sv ∈ Γξ to maximize L(w − αgt w; Sv), we do first order Taylor expansion

Convergence of this alternative iteration

​

​

​

They only compute gradient w.r.t. parameters of the final layer of learner

L2 - Normalisation

Experiments

​

Model : Resnet18 , Resnet 50 as feature extractor ( fe ) , fn (L2 norm) , fc (classification layer with marginal softmax(s,m)

Data : PACS - art painting (A), cartoon (C), photo (P), sketch (S) sharing 7 classes

Office-home - Art (Ar), Clipart (Cl), Product (Pr), Real World (Rw), sharing 65 classes

CIFAR10 - single domain

Setups :

1. MULTI SOURCE WITH DOMAIN SHIFT (MSDS)
2. Single Source with Domain Shift (SSDS)
3. Same Source and Target Domain (SSTD)

1.MSDS

In the setting of MSDS,

• leave-one-domain-out cross validation, i.e., training on three domains and testing on the remaining unseen domain, on PACS and Office-Home.

Note that the domain labels are not used during training.

They adopt ResNet18 and ResNet50 pre-trained on ImageNet as feature extractor fe

Observations

On PACS, DFAS significantly outperforms Baseline in almost all tasks except when P is taken as target domain.

Note that, in the task that domain S, of which the style is extremely different from rest three domains, is target domain, DFAS boosts the accuracy significantly .

This indicates that our method can generalize well when domain shift is large

2.SSDS

We conduct this type of experiment on PACS based on ResNet18.

In this experiment, we train learner on one domain and test on each of the rest three domains, resulting in total 12 tasks.

observations

This method of DFAS outperforms Baseline and SagNet.

It outperforms Baseline in 10 tasks among all 12 tasks.

​

Effectiveness on single source domain training is observed when the unseen target domain is different from source domain.

SSTD

Applied DG method to the common recognition task that the training and test data which are from a same domain, i.e., SSTD, on CIFAR-10 dataset.

To investigate the effect of training size, we sample different sizes of training data from the provided training set (i.e., source domain)

Observations

DFAS outperforms Baseline and compared methods in all different numbers of training data.

In general, the performance boost is larger when the number of training data is smaller.

This may be because the learner is more possible to be overfitting when the training size is smaller and DFAS is designed to extract better generalizable features.

Ablation Study

From these experimentations :

1. Effectiveness of L2
2. Effectiveness of adversarial split
3. Effectiveness of adversarial over random and domain label

Were observed

Additional Details

1. MITIGATING GRADIENT EXPLOSION BY L2-NORMALIZATION

​

2) EFFECT OF HYPER-PARAMETERS

Note : Theoretical analysis to prove their method can found in paper.

T-SNE visualisations of fe

Takeaways

1. Generally as adverersial images are used to make the model robust to changes, here adversarial splitting is used to make the model generalise in different domain scenarios
2. This adversarial splitting model uses domain free setting to obtain better learner which is trained used the meta task training , where the train / val split has the max loss (min max optimisation).
3. This type of training mechanism helped the model learn domain agnostic and try to achieve DG.

Open questions

​

1. Are the self supervised learning tasks that are starting to become mainstream better than this meta learning approach?
2. Are there any other possibilities other than domain shift when L(Sv) is max

​

Discussion