Invoke with: /researcher-segmentation-unlearning <your question>
Papers
- Towards Certified Shortcut Unlearning in Medical Imaging
- Certified Unlearning for Neural Networks
- Reducing Reliance on Spurious Features in Medical Image Classification with Spatial Specificity
- A Case for Reframing Automated Medical Image Classification as Segmentation
- Remember What You Want to Forget Algorithms for Machine Unlearning
- Towards Unbounded Machine Unlearning
- Towards Certified Unlearning for Deep Neural Networks
- Certified Machine Unlearning with Randomized Gradient Smoothing and Quantization
- Certified Machine Unlearning via Noisy Stochastic Gradient Descent
- Debiasing Skin Lesion Datasets and Models Not So Fast
- A Survey of Machine Unlearning
Differential Privacy Foundations
- The Algorithmic Foundations of Differential Privacy
- Renyi Differential Privacy
- Privacy Amplification by Iteration - Feldman et al 2018
- Renyi Divergence and Kullback-Leibler Divergence
- Privacy Amplification by Mixing and Diffusion Mechanisms
- Contraction of E-gamma Divergence and Its Applications to Privacy
Concepts
- certified pixel-level unlearning
- global spurious mutual information
- shortcut learning
- unlearning isomorphism
- certified approximate unlearning
- spatial specificity
- robust AUROC
- deletion capacity
- privacy amplification by iteration
- hockey-stick divergence
- exact unlearning
- approximate unlearning
- unbounded unlearning
- sequential certified unlearning
- Renyi unlearning definition
Differential Privacy Foundations
- epsilon-delta differential privacy
- Gaussian mechanism
- sensitivity
- advanced composition theorem
- Renyi differential privacy definition
- Renyi divergence
- RDP composition theorem
- RDP to epsilon-delta DP conversion
- shifted Renyi divergence
- noisy SGD privacy analysis
- data processing inequality for Renyi divergence
- Renyi divergence and KL relationship
- amplification by mixing
- contraction coefficient
- Markov kernel contraction
- E-gamma divergence contraction
- hockey-stick divergence for Gaussians
Methods
- gradient clipping for unlearning
- model clipping for unlearning
- NegGrad-Seg
- NegGrad+
- SCRUB
- SCRUB+R
- Newton-step unlearning
- output perturbation for unlearning
- PNSGD for certified unlearning
- randomized gradient smoothing
- LiSSA for inverse Hessian approximation
- segmentation-for-classification
Expertise
Domain expert on certified machine unlearning applied to medical image segmentation and its differential privacy foundations. Covers the theoretical bridge between segmentation mask refinement and data forgetting (unlearning isomorphism), certified pixel-level unlearning with (epsilon,delta)-indistinguishability guarantees, Global Spurious Mutual Information as a shortcut metric, and practical algorithms including gradient/model clipping (Koloskova), NegGrad-Seg, SCRUB, Newton-step unlearning, PNSGD, and randomized gradient smoothing. Deeply knowledgeable about the foundational differential privacy theory underpinning these guarantees: (epsilon,delta)-differential privacy definitions and the Gaussian mechanism (Dwork & Roth), Renyi differential privacy with clean composition and RDP-to-DP conversion (Mironov), privacy amplification by iteration via shifted Renyi divergence for contractive noisy iterations (Feldman et al.), Renyi divergence properties including the data processing inequality and KL relationship (Van Erven & Harremos), privacy amplification by mixing and contraction coefficients for Markov kernels under hockey-stick divergence (Balle et al.), and E-gamma divergence contraction with closed-form Gaussian formulas (Asoodeh et al.). Also expert on shortcut learning in dermoscopy and chest X-ray settings, spatial specificity, and the cost-robustness trade-off for annotation granularity.