First Step Is Alpha

Adam with the paper's defaults: α=0.001\alpha = 0.001, β1=0.9\beta_1 = 0.9, β2=0.999\beta_2 = 0.999, ϵ=10−8\epsilon = 10^{-8}, and m0=v0=0m_0 = v_0 = 0.

At step t=1t = 1 a particular parameter sees gradient g1=0.1g_1 = 0.1.

m1=β1m0+(1−β1)g1,v1=β2v0+(1−β2)g12m_1 = \beta_1 m_0 + (1-\beta_1) g_1, \qquad v_1 = \beta_2 v_0 + (1-\beta_2) g_1^2

m^1=m11−β1 1,v^1=v11−β2 1,Δθ=−α m^1v^1+ϵ\hat{m}_1 = \frac{m_1}{1-\beta_1^{,1}}, \qquad \hat{v}_1 = \frac{v_1}{1-\beta_2^{,1}}, \qquad \Delta\theta = -\alpha,\frac{\hat{m}_1}{\sqrt{\hat{v}_1}+\epsilon}

How large is ∣Δθ∣|\Delta\theta|?

A

About 0.10.1.

B

About 0.0010.001, that is α\alpha itself.

C

About 0.00320.0032.

D

About 10−510^{-5}.