We present a few hypotheses as to why data poisoning is so tricky to remove. Some hypotheses include that data poisoning moves the model parameters a lot, or in a subspace orthogonal to the clean data. Would love to see more exploration here! 6/n
@thegautamkamath
-
MUL Method Mitigates Data Poisoning Effects in Training
By
–
But MUL has also been proposed as a method to remove the effect of data poisoning. Such effects are *indirect* — they manifest in other datapoints besides the specific training point. 4/n
-
Machine Unlearning: Removing Training Data Influence Efficiently
By
–
MUL is pretty well studied by now: given a trained model, can we "remove" all influence of some points in the training set? Naive way: retrain from scratch without those points. But this is slow. A number of faster methods have been proposed. 2/n
-
Evaluating Machine Unlearning with Membership Inference Attacks
By
–
The most common way to evaluate an MUL method is using "membership inference attacks." These try to directly test whether the "unlearned" point was in the dataset used to train the model or not. 3/n
-
Machine Unlearning Fails Against Data Poisoning Attacks
By
–
New paper: Machine Unlearning Fails to Remove Data Poisoning Attacks, ft @MartinPawelczyk
, @jimmy_di98
, @ayush_sekhari
, @SethInternet
. Title says it all: current approaches for machine unlearning (MUL) are not effective at removing the effect of data poisoning attacks. 1/n -
NeurIPS Papers Show Low Impact Despite Acceptance
By
–
Most accepted NeurIPS papers are no to low impact
-
ACM-SIAM Discrete Algorithms Symposium Submission Deadline Approaching
By
–
Everyone submitting to the ACM-SIAM Symposium on Discrete Algorithms right now #SIAMDA25pic.twitter.com/lSD6paiyz7
— Gautam Kamath (@thegautamkamath) 6 juillet 2024Everyone submitting to the ACM-SIAM Symposium on Discrete Algorithms right now #SIAMDA25
-
University Teaching Flexibility and Parental Leave Benefits
By
–
At @UWCheritonCS we have a lot of flexibility. You choose which terms you teach (we have three each year, due to a co-op program), as long as you keep a reasonable balance. Parental leave gives you some free teaching credits towards your balance.
-
ML Security: Why Attacking Systems Validates Their Safety
By
–
*Must read* for anyone interested in ML security, by Nicholas Carlini. Attacks are the only way we know whether or not a purportedly secure system actually is. Moreover, I consider personal attacks like this unacceptable in my research communities. https://
nicholas.carlini.com/writing/2024/w
hy-i-attack.html
… -
Luca Trevisan’s Final Talk on Spectral Graph Theory
By
–
In case you missed it, Luca Trevisan's final talk was delivered today by @praveshkkothari and Salil Vadhan. Part 1 is on spectral graph theory, and part 2 is about his experiences as an outsider. Things start around 22:30. https://
youtube.com/live/FdP_n_x0a
1k?si=5BNaFKS9cFvjwzZo
…