Kwangjin Oh
@heroide
ID: 99101594
24-12-2009 13:15:02
14 Tweet
42 Followers
166 Following
μλ§κ° ννλ¬μ€μμ λΉμ²¨λμ΄μ λ€κ³ μ¨ νΈλμ νκΈ°μΈλ°γ γ γ γ γ γ γ μμΉ¨μ λ΄κ° 건μ μ§ κ½μμ λλ Έλλ μ΄λ° γ γ γ γ γ γ γ γ γ γ μ΄κ±° μ΄λ»κ² λ€κ³ λ€λλΌκ³ γ γ γ γ γ γ γ γ γ γ μ€λ ν루μ μμμ§λ’°κ° λμ΄λ²λ ΈλΌγ γ γ γ γ γ
Re-parameterize all the things! If you want to backprop thru samples from mixture, truncated, Gamma, Beta, Dirichlet, Student-t, or von Mises distributions, this paper has gotchu covered. Also faster than RSVI, another general reparameterization trick based on rejection sampling x.com/BrundageBot/stβ¦
Implicit Reparameterization Gradients by myself, Shakir Mohamed and Andriy Mnih: arxiv.org/abs/1805.08498. An extension of reparameterization for many distributions, including Gamma, Beta, Dirichlet and von Mises, that is faster and easier to use than the previous methods.
