#ConformalPrediction #VisionTransformers #SkinLesionClassification #UncertaintyQuantification #MedicalAI #DomainAdaptation #AIResearch

Domain Adaptation of Base Models + ShadowdarkQA Bench

https://gygaxtest.com/posts/continued_pretraining_for-rules/

#HackerNews #DomainAdaptation #ShadowdarkQA #BaseModels #AIResearch #MachineLearning

I am reading several articles about domain adaptation for medical imagery, especially for #MRI. It seems people are stuck in certain patterns of ideas. They train multiple models for different domains and combine, they condition the model with some known imaging parameters or subject type parameters, ...

I did a quick literature search and didn't find much else than those approaches. Now with #DeepLearning we can do much better actually.

What I see missing is like just give the neural network one example calibration image taken with the device. Take it of a banana for all I care, as long as it's roughly of a similar subject. Now you have a cheap and clear example image from each device at each #hospital. You must already have such, would be surprising if you didn't have any image of a healthy subject from an arbitrary device.

Now instead of any smartypants stuff, just give this image to a #CNN which conditions the task network proper by providing it with domain calibration conditioning. Train it jointly with many different segmentation tasks.

You only need to be slightly careful so that you don't accidentally make the #DomainAdaptation network learn specific hospital identities, but segmentation tasks should be relatively robust against that risk. Diagnosis classification tasks wouldn't be.

Anyhow, I'm just saying there are better ways to do these things if you let go of the established and calcified academic thinking which was a product of the time before deep learning.

Because of different populations, and observation methods, a trained #MachineLearning predictor has to deal with out of distribution prediction. #CovariateShift #DomainAdaptation

Hence it is hard to say what the performance is going to be on any new data. #Medicine #Health

https://bmcmedicine.biomedcentral.com/articles/10.1186/s12916-023-02779-w

Domain-adaptation Fine-tuning of Foundation Models in Amazon SageMaker JumpStart on Financial data https://triangleagency.co.uk/domain-adaptation-fine-tuning-of-foundation-models-in-amazon-sagemaker-jumpstart-on-financial-data/?utm_source=dlvr.it&utm_medium=mastodon #TheTriangleAgencyNews #Amazon #Data #Domainadaptation