<ul><li>Study on convergence properties and escape dynamics of Stochastic Gradient Descent (SGD) in one-dimensional landscapes with infinite- and finite-variance noise.</li><li>Focus on identifying time scales for SGD to move from initial point to local minimum in the same basin.</li><li>SGD converges to basin's minimum unless initial point is too close to a local maximum, leading to lingering in its neighborhood.</li><li>Results show SGD does not remain stuck near a 'sharp' maximum and provide estimates on reaching neighboring minima, influenced by noise characteristics and function geometry.</li></ul>

Convergence, Sticking and Escape: Stochastic Dynamics Near Critical Points in SGD

Discover more