Peter Gehler retweeted
Excited to share that my internship work @AmazonScience was accepted at WACV 2023!
Do you want to adapt your computer vision model to a new distribution while using ~5-10x less data? Check out TeST: a 2 stage Test-Time adaptation model!
@pegehler @FrancescoLocat8 Bernt Schiele
Peter Gehler retweeted
Can #NeuralNetworks generalize to factors of variation present in their training data?
We investigated this with a large-scale study and showed limited generalization capabilities.
Benchmark: github.com/bethgelab/InDomai…
Paper: arxiv.org/pdf/2107.08221.pdf
#MachineLearning #AI
[1/6]
Peter Gehler retweeted
What's the role of pre-trained representations for RL and how important is this for OOD generalization? We attempted to find some answers on that!
“Representation Learning for Out-of-Distribution Generalization in Reinforcement Learning”
arxiv.org/abs/2107.05686
[1/5]
Peter Gehler retweeted
High quality in industrial production requires accurate visual inspection to remove defective parts.
We develop a method, trained on normal images only, which achieves less than half the error of previous methods. @awscloud @AmazonScience
Link: arxiv.org/abs/2106.08265v1
[1/4]
Peter Gehler retweeted
The State-Of-The-Art in AI changes all the time. What do you do if you made predictions on some data and someone releases a better model? We develop a probabilistic approach for backward-compatible prediction updates. @awscloud @AmazonScience
draft: arxiv.org/pdf/2107.01057.pdf
