SSU-LSF
The first machine unlearning framework targeting geoscience Mamba-based structured state space models. It achieves targeted removal of non-stationary confounding event influences through a dedicated EKFac influence function, a spectral-radius-weighted inflection threshold method, and Hessian-free projected gradient ascent.
Timeline (last 90 days)
SSU-LSF claims to be the first machine unlearning framework specifically designed for geoscience Mamba-class state space models.
SSU-LSF uses dedicated EKFac influence functions for the Mamba state matrix to quantify the contribution of events to parameters via closed-form matrix exponential gradients.
SSU-LSF employs a spectral radius-weighted inflection point threshold method to localize confounding footprints in time.
SSU-LSF uses Hessian-free projected gradient ascent to perform the forgetting operation within a KL divergence trust region, supplemented by spatial total variation regularization.
Proposition 1 of the paper proves that residual confounding is bounded by a bound related to the length of the contamination window, with a longer contamination window yielding a larger upper bound on residual bias.
The confounding reduction rates of SSU-LSF on the three benchmarks CropHarvest, NDVI-LST, and ERA5 are 0.773, 0.821, and 0.859, respectively.
On ERA5, the worst-case RMSE degradation after SSU-LSF forgetting is 4.2%.
SSU-LSF converges within 3 to 5 epochs, and the GPU cost per forgetting request is 8.4 times lower than full retraining.
The authors evaluated against eleven baseline methods on three heterogeneous benchmark datasets.
SSU-LSF is an approximate unlearning method that, targeting the characteristic of the Mamba state transition matrix appearing in matrix exponential form, computes gradients using the closed-form solution of the Fréchet derivative of the matrix exponential.
1 more timeline events
All Facts (11)
SSU-LSF claims to be the first machine unlearning framework specifically designed for geoscience Mamba-class state space models.
50%UnverifiedSSU-LSF uses dedicated EKFac influence functions for the Mamba state matrix to quantify the contribution of events to parameters via closed-form matrix exponential gradients.
50%UnverifiedSSU-LSF employs a spectral radius-weighted inflection point threshold method to localize confounding footprints in time.
50%UnverifiedSSU-LSF uses Hessian-free projected gradient ascent to perform the forgetting operation within a KL divergence trust region, supplemented by spatial total variation regularization.
50%UnverifiedProposition 1 of the paper proves that residual confounding is bounded by a bound related to the length of the contamination window, with a longer contamination window yielding a larger upper bound on residual bias.
50%UnverifiedThe confounding reduction rates of SSU-LSF on the three benchmarks CropHarvest, NDVI-LST, and ERA5 are 0.773, 0.821, and 0.859, respectively.
50%UnverifiedOn ERA5, the worst-case RMSE degradation after SSU-LSF forgetting is 4.2%.
50%UnverifiedSSU-LSF converges within 3 to 5 epochs, and the GPU cost per forgetting request is 8.4 times lower than full retraining.
50%UnverifiedThe authors evaluated against eleven baseline methods on three heterogeneous benchmark datasets.
50%UnverifiedSSU-LSF is an approximate unlearning method that, targeting the characteristic of the Mamba state transition matrix appearing in matrix exponential form, computes gradients using the closed-form solution of the Fréchet derivative of the matrix exponential.
50%UnverifiedThe confounding reduction rate of SSU-LSF ranges between 0.77 and 0.86, meaning a considerable proportion of confounding effects remain uneliminated, and the method is still a preprint paper.
50%