Contrasting Multiple Representations with the Multi-Marginal Matching Gap
AuthorsZoe Piran, Michal Klein, James Thornton, Marco Cuturi Cameto
Contrasting Multiple Representations with the Multi-Marginal Matching Gap
AuthorsZoe Piran, Michal Klein, James Thornton, Marco Cuturi Cameto
Learning meaningful representations of complex objects that can be seen through multiple () views or modalities is a core task in machine learning. Existing methods use losses originally intended for paired views, and extend them to views, either by instantiating loss-pairs, or by using reduced embeddings, following a strategy. We propose the multi-marginal matching gap (M3G), a loss that borrows tools from multi-marginal optimal transport (MM-OT) theory to simultaneously incorporate all views. Given a batch of points, each seen as a -tuple of views subsequently transformed into embeddings, our loss contrasts the cost of matching these ground-truth -tuples with the MM-OT polymatching cost, which seeks optimally arranged -tuples chosen within these vectors. While the exponential complexity ) of the MM-OT problem may seem daunting, we show in experiments that a suitable generalization of the Sinkhorn algorithm for that problem can scale to, e.g., views using mini-batches of size . Our experiments demonstrate improved performance over multiview extensions of pairwise losses, for both self-supervised and multimodal tasks.
Private Stochastic Convex Optimization with Heavy Tails: Near-Optimality from Simple Reductions
November 20, 2024research area Methods and Algorithms, research area Privacyconference NeurIPS
We study the problem of differentially private stochastic convex optimization (DP-SCO) with heavy-tailed gradients, where we assume a -moment bound on the Lipschitz constants of sample functions, rather than a uniform bound. We propose a new reduction-based approach that enables us to obtain the first optimal rates (up to logarithmic factors) in the heavy-tailed setting, achieving error $G_2 \cdot \frac 1 {\sqrt n} + G_k \cdot…
Private Frequency Estimation via Projective Geometry
July 11, 2022research area Methods and Algorithms, research area Privacyconference ICML
In this work, we propose a new algorithm ProjectiveGeometryResponse (PGR) for locally differentially private (LDP) frequency estimation. For a universe size of and with users, our -LDP algorithm has communication cost bits in the private coin setting and in the public coin setting, and has computation cost for the server to approximately…