y, but have yet to match autoregressive models on density estimation benchmarks. In this paper we make several technical contributions that allow diffusion models to challenge the dominance of autoregressive models in this domain.
문제제기 : 아직 autoregressive 모델과 일치하지 않음. 맞추게하겠음
The distribution of latent variable zt conditioned on x, for any t ∈ [0, 1] is given by:
where αt and σ 2 t are strictly positive scalar-valued functions of t. Furthermore, let us define the signal-to-noise ratio (SNR):
알파와 분산은 t의 positive scalar-valued 함수로 쓰임. 저자는 signal-to-noise ratio(SNR)을 정의했음.
We assume that the SNR(t) is strictly monotonically decreasing in time, i.e. that SNR(t) < SNR(s) for any t > s. This formalizes the notion that the zt is increasingly noisy as we go forward in time.
저자는 SNR(t)는 시간에 단조롭게 하향함. 이 공식은 시간이 앞으로 갈수록 노이즈가 증가한다는 의미임.
In previous work, the noise schedule has a fixed form (see Appendix H, Fig. 4a). In contrast, we learn this schedule through the parameterization
where γη(t) is a monotonic neural network with parameters η, as detailed in Appendix H
noise schedule은 고정되어있음. Fig4보면 알수 있음. 대조적으로 저자는 이 schedule을 학습할 수 있다. γη(t)는 monotonic neural network임.
It is straightforward to verify that α 2 t and SNR(t), as a function of γη(t), then simplify to:
위와 같인 간단하게 표현할 수 있음.