Current browse context:
cs.LG
Change to browse by:
References & Citations
Computer Science > Machine Learning
Title: Sample Efficient Learning of Factored Embeddings of Tensor Fields
(Submitted on 1 Sep 2022 (v1), last revised 9 Mar 2024 (this version, v2))
Abstract: Data tensors of orders 2 and greater are now routinely being generated. These data collections are increasingly huge and growing. Many scientific and medical data tensors are tensor fields (e.g., images, videos, geographic data) in which the spatial neighborhood contains important information. Directly accessing such large data tensor collections for information has become increasingly prohibitive. We learn approximate full-rank and compact tensor sketches with decompositive representations providing compact space, time and spectral embeddings of tensor fields. All information querying and post-processing on the original tensor field can now be achieved more efficiently and with customizable accuracy as they are performed on these compact factored sketches in latent generative space. We produce optimal rank-r sketchy Tucker decomposition of arbitrary order data tensors by building compact factor matrices from a sample-efficient sub-sampling of tensor slices. Our sample efficient policy is learned via an adaptable stochastic Thompson sampling using Dirichlet distributions with conjugate priors.
Submission history
From: Taemin Heo [view email][v1] Thu, 1 Sep 2022 11:32:00 GMT (6323kb,D)
[v2] Sat, 9 Mar 2024 03:16:12 GMT (9110kb,D)
Link back to: arXiv, form interface, contact.