This blog post is in the works.
Why should you care about Domain Generalization (DG)?
Developing robust models that are able to generalize to out-of-distribution (OOD) data has become a popular topic for many machine learning researchers to work on in order to improve the generalizeability of applied machine learning.
This blog aims at providing an introduction to domain generalization by:
- Defining the jargon
- Providing a clear explanation on the problem setting
- Highlighting seminal work and important domain generalization methods
- Provide up to date resources (talks/papers/codebases)
Defining the jargon
In domain generalization work, as well as in machine learning more broadly, there are specific words that have not been consistently defined. In this blog I have tried to provide the most intuitive explanation of terms that I believe are being used in a hand wavy way and I have vetted my definitions by experts in this work.
Here are a list of terms I plan to clarify:
- Spurious Correlations
Make a clear distinction between the following work (ie. definition and maybe a seminal paper in that field and relevent resource)
- Domain generalization
- Meta learning
- Multi task learning
- Distributional robustness
- Adversarial Robustness
- Out of distribution detection
- Domain adaptation
- Zero-shot learning
- Few-shot learning
- Transfer learning
- Negative Transfer Learning
THIS IS AN EXAMPLE ON HOW I WANT TO DEFINE SOME OF THE TERMINOLOGY THAT IS COMMONLY USED.
Whose model is more robust?
Now that we have the data, let’s just be clear on the distributions:
So who was right?
Your model is more robust! Although I achieved higher accuracy on the data with a distribution shift, my model saw a 16% drop in performance whereas yours only saw an 8% drop. The lower the drop in performance between distributions, the more robust a model is.
ICP (Invariant Causal Prediction) for Nonlinear models
Generalizing from Several Related Classification Tasks to a New Unlabeled Sample
Causal Inference Using Invariant Prediction: Indentification and Confidence Intervals
Mathematical formulation of the problem
Visual intuitive and laymen explanation of the problem
Seminal Work and DG Methods
ICP and Non-Linear ICP
Invariant Risk Minimization (IRM) & IRMV2
Discuss limitations and method formulation and maybe relevant papers for each of these
Robust Optimization Methods
DG & Algorithmic Fairness
DG for Decomposition
DG for Meta Learning
Self-supervised Contrastive Regularization (SelfReg)
Probably talk about WILDS here and a lot of Pearcy Liang’s work
DG and Causal Inference
DG Gradient Alignment
DG Theory on Sample Complexity & PAC Learning
DG for Computer Vision and Data Augmentation
Resources to get started
If you are interested in getting involved in domain generalization work, I would recommend first cleaning up your understanding on causal inference and distributionally robust optimization. For lack a better word, these can be considered your “pre-requisities”.
General causal inference content
Papers to motivate causal inference in domain generalization
- Here is a great list of paper
Distributionally Robust Optimization
General causal inference content
Great researchers to follow on Twitter
List a good list of people doing DG work like Chealsea Finn or Piearcy Liang