Understanding Over-parametrization Through Matrix Sensing

Channel:
Microsoft Research
Subscribers:
351,000
Published on ● Video Link: https://www.youtube.com/watch?v=D4HLmgGwt1o


Duration: 1:17:49
1,489 views
13

We study the problem of recovering a low-rank matrix from linear measurements using an over-parameterized model. We show that the gradient descent process on the square loss function, starting from a small initialization, can converge to the ground truth matrix. Although the total number of observations is much smaller than the total number of parameters.

See more at https://www.microsoft.com/en-us/research/video/understanding-over-parametrization-through-matrix-sensing/



Other Videos By Microsoft Research


2018-02-06Project Belgrade, User Stories - Katrine
2018-02-06Project Belgrade, User Stories - Steven
2018-02-04Toward Scalable Fontlings - July 24, 2006
2018-02-04Kid Tab Demo - Part 1
2018-02-03Kid Tab Demo - Part 2
2018-02-02Hand Representations for Typing in VR VR 2018
2018-02-02Text Entry in Immersive HMDisplay based VR using Physical and Touch Keyboards
2018-01-31Data Wrangling using Programming by Examples
2018-01-26Experimentally Reducing Partisan Incivility on Twitter
2018-01-25Visual Understanding in Natural Language
2018-01-25Understanding Over-parametrization Through Matrix Sensing
2018-01-24Sustainable Computing at Scale for Smart Cities
2018-01-24Transport Modelling and user Behavior in Urban Systems
2018-01-24Frugal Innovations for Rural Areas - Panel Discussion
2018-01-22A peek into Digital Green
2018-01-15Fiat Cryptography: Automatic Correct-by-Construction Generation of Low-Level Cryptographic Code
2018-01-15Building Scale VR: Creating Indoor 3D Maps & its Application to Simulation of Disaster Situations
2018-01-12Digital Society: Panel Discussion
2018-01-11Lokavidya
2018-01-10CRISPR.ML - Machine learning meets gene editing
2017-12-22Hitting the Nail on the Head: Interdisciplinary Research in Computer Networking


Tags:
microsoft research
low-rank matrix
over-parameterized model