2329. DISM: A Distributed Incremental Subgradient Method for Weakly Supervised Skin Segmentation
Invited abstract in session MD-50: Variational analysis and equilibria, stream Variational analysis, equilibria and nonsmooth optimization.
Monday, 14:30-16:00Room: Parkinson B11
Authors (first author is the speaker)
| 1. | Narges Araboljadidi
|
| Department of Mathematics and Physics, University of Campania ”Luigi Vanvitelli” | |
| 2. | Valentina De Simone
|
| Mathematics and Physics, University of Campania "L. Vanvitelli" |
Abstract
We present a novel Distributed Incremental Subgradient Method (DISM) for Multiple Instance Learning (MIL) applied to skin segmentation. Traditional approaches require labor-intensive pixel-level annotations, while our weakly supervised framework uses only image-level labels, treating images as bags of pixels with binary indicators of skin presence. Our method optimizes the resultant non-smooth, non-convex objective function with an efficient distributed architecture and reduced-synchronization strategies that prevent parallel processing overhead with retained convergence properties. Experiments with the UCI skin segmentation dataset show our method achieving significantly higher accuracy (79.25\%) compared to the MATLAB toolbox method with sequential minimal optimization with SVM (59.81\%), with 50.3\% computational time saved (from 17.38 to 8.64 seconds). This contribution implies how distributed MIL can effectively leverage weak labels to counteract annotation load with greater scalability for large-scale segmentation problems with high accuracy and computational savings.
Keywords
- Machine Learning
- Non-smooth Optimization
- Programming, Nonlinear
Status: accepted
Back to the list of papers