Steepest Descent

Photo by Davide Foti on Unsplash

“Follow the direction of steepest descent”
is what they told me,
but my loss landscape is not convex,
and I may have only found
a local optimum.

“Stochasticity will help”
they said,
but my life is already noisy enough.

“Adapt your moments”
they suggested,
but I never could converge.

My gradients have vanished.

I guess I have too many layers to be useful.

Tomo Lazovich 2019