Nick Jia

Researching Trustworthy Machine Learning advised by @NicolasPapernot
PhD student at @UofT and @VectorInst

This is my only X account

[Pratyush Maini @pratyushmaini]

1/We've nailed a framework to reliably detect if an LLM was trained on your dataset: LLM Dataset Inference.

After over a year of thinking of writing about how hard this is, we had a breakthrough that made me quite literally jump from my seat!

📝: https://t.co/B9jvMsoNGW Long🧵

[Varun Chandrasekaran @VarunChandrase3]

I am looking to recruit students at UIUC (deadline: Dec 15, 2022). More details (about me/my research interests) are on my webpage. RTs are appreciated!

[Nick Jia @NickJia5]

@usmannk @NicolasPapernot @HONGYUC12266081 @jonasguan @AliShahinShams1 @UofT @VectorInst @TorontoSRI @turinginst it is possible that models with different weights may have similar functionality. Therefore, I would suggest using our method for this case since it is designed to compare the global behavior/functionality of ML models.

[Nick Jia @NickJia5]

@usmannk @NicolasPapernot @HONGYUC12266081 @jonasguan @AliShahinShams1 @UofT @VectorInst @TorontoSRI @turinginst DNNs are usually overparameterized. In other words, since the weight space is usually much higher-dimensional than the space of mapping from inputs to outputs which characterizes the functionality of the models,

[Nick Jia @NickJia5]

@usmannk @NicolasPapernot @HONGYUC12266081 @jonasguan @AliShahinShams1 @UofT @VectorInst @TorontoSRI @turinginst If I understand correctly, your question is how to compare two models when they have the same architecture but are trained with different initialization, random seed, hyperparameters, and/or even datasets. In this case, I agree with you that direct weight similarity is hard.

[Nick Jia @NickJia5]

Paper: https://t.co/fCndYrqsbp
Poster: https://t.co/CxzMbv4Ful

[Nick Jia @NickJia5]

How to measure the similarity between 2 DNNs with different architectures? We propose to approximate DNNs' global behaviors using their local behaviors captured by model explanation techniques such as LIME, and compute the distance on the latter. #ICLR2022

[Nick Jia @NickJia5]

We have also showed that such a distance metric might be used to verify the correctness of machine unlearning.

[Nick Jia @NickJia5]

We have also shown that such a distance metric might be used to verify the correctness of machine unlearning.

[Nick Jia @NickJia5]

@NicolasPapernot @EmmyFangCY @DeepMind @UofT @VectorInst @BoWang87 Congrats! @EmmyFangCY

[Nick Jia @NickJia5]

@NicolasPapernot @VectorInst @uoftengineering Thank you!