Views Navigation

Event Views Navigation

High-dimensional limit theorems for Stochastic Gradient Descent: effective dynamics and critical scaling

Gérard Ben Arous (NYU Courant)
E18-304

Abstract: We study the scaling limits of stochastic gradient descent (SGD) with constant step-size in the high-dimensional regime. We prove limit theorems for the trajectories of summary statistics (i.e., finite-dimensional functions) of SGD as the dimension goes to infinity. Our approach allows one to choose the summary statistics that are tracked, the initialization, and the step-size. It yields both ballistic (ODE) and diffusive (SDE) limits, with the limit depending dramatically on the former choices. Interestingly, we find a critical scaling…

Find out more »


© MIT Institute for Data, Systems, and Society | 77 Massachusetts Avenue | Cambridge, MA 02139-4307 | 617-253-1764 |
      
Accessibility