Train your own embedding model
Backprop, by hand·2 min

Why finite differences?

0 / 1

Every reference gradient in this chapter is cross-checked against central finite differences before it ships. One quick question on why that matters.

Knowledge check

We validate the reference gradients with central finite differences (L(θ+ε) − L(θ−ε)) / 2ε. Why bother, instead of just trusting the analytic derivation?