Reference
2011.13456.pdf (arxiv.org) : good reference for Diffusion to SDE from Stanford
PII: 0304-4149(82)90051-5 (core.ac.uk) @andersonReversetimeDiffusion1982 : 專門針對 reverse time SDE equations.
http://pordlabs.ucsd.edu/pcessi/theory2019/gardiner_ito_calculus.pdf
Ito equations
Introduction
全部的重點就是下圖。用一個神經網路逼近 score function! 因爲我們不知道 $p(x)$
Background
Method 1: denoising Score Matching with Langevin Dynamics (SMLD)
主要是利用 neural network 去近似 p(x) 的 score function, i.e. gradient of log likelihood.
Method 2: denoising diffusion probabilistic model (DDPM)
SDE and Reverse SDE
Forward SDE
Backward SDE
DDPM (Variance Preserving SDE)
SMLD (Variance Exploding SDE)
Sub-Variance Preserving SDE
