Read on Twitter
The CleverHans Lab's work on machine unlearning was accepted to @IEEESSP 
Excited to share the paper, blog post, code. I am very proud of the work that my students accomplished @LucasBourtoule @VarunChandrase3 @Chris_Choquette @NickJia5 @alkae_t @zhangbaiwu joint w/ @david_lie
Excited to share the paper, blog post, code. I am very proud of the work that my students accomplished @LucasBourtoule @VarunChandrase3 @Chris_Choquette @NickJia5 @alkae_t @zhangbaiwu joint w/ @david_lie
What should an organization do with its machine learning models when users request that their data be deleted, after they previously agreed to share it? This is the problem we tackle. In general, it is very difficult to provably unlearn a data point.
This is because to erase its contributions we would first have to calculate them. We develop an unlearning algorithm that outputs the same distribution of models than one would obtain by retraining without the points to be unlearned.
We introduce 2 optimizations to the naive strategy of retraining from scratch:
(1) Sharding speeds up retraining by reducing the total number of data points that are affected by an unlearning request
(1) Sharding speeds up retraining by reducing the total number of data points that are affected by an unlearning request
(2) Slicing reorders data during training to enable us to save useful intermediate checkpoints of the model. When an unlearning request is received, this allows us to retrain the model from a checkpoint rather than a random initialization.
This is the first paper that we wrote as a lab!
I wanted to thank:
- the reviewers and especially our point of contact Henry
Corrigan-Gibbs for constructive feedback throughout the revision process.
I wanted to thank:
- the reviewers and especially our point of contact Henry
Corrigan-Gibbs for constructive feedback throughout the revision process.
- the staff @UofT @VectorInst who support us every day. Among other things, students are lucky to have access to a large cluster of GPUs @VectorInst supported by a terrific scientific computing team.
- @CIFAR_News for providing us with "no strings attached" funding giving us the freedom to pursue ideas like this. All they ask for is 25mn of my time once a year: this makes a big difference in being able to spend more time on research rather than managing funding
and (less seriously) @TimHortons for providing us with enough donuts to power through the last few hours before the submission deadline
