Korean, Edit

Flow Algorithm

Recommended post: 【Algorithm】 Algorithms·Machine Learning Table of Contents


1. Overview

2. Type 1. CNF

3. Type 2. DOT

4. Type 3. Flow matching



1. Overview

⑴ Streamline: for vector field f(x, t) and position x(t)


스크린샷 2026-07-04 오후 1 11 31

스크린샷 2026-07-04 오후 1 12 06

Figure 1. Example of a streamline


Continuity equation: by the law of conservation of mass, the density ρ(x, t) must satisfy the following


스크린샷 2026-07-04 오후 1 12 29


⑶ Flow algorithm


스크린샷 2026-07-04 오후 1 13 00


⑷ The uniqueness of f becomes an issue, and if uniqueness is satisfied, it becomes the domain of causal inference (ref)



2. Type 1. CNF

⑴ Continuous normalizing flow (CNF): learning a complex distribution ρ(x, t1) from a simple distribution ρ(x, t0), as in a diffusion model (Grathwohl et al., 2019)


스크린샷 2026-07-04 오후 1 14 13


⑵ In actual papers, it is expressed more specifically as follows


스크린샷 2026-07-04 오후 1 14 39



3. Type 2. DOT

Optimal transport theory(optimal transport)

⑵ Dynamic optimal transport (DOT): constructs a continuous path that minimizes the cost of the trajectory connecting two time points t0 and t1. Such a cost is called the L2 Wasserstein distance


스크린샷 2026-07-04 오후 1 15 06


⑶ In the above equation, ρ ㅣfㅣ2 ≃ mv2 = 2K, which is closely related to energy.

Example 1. TrajectoryNet implements DOT using CNF.

Example 2. DeepRUOTv2



4. Type 3. Flow matching

⑴ Flow matching: although its objective is the same as CNF, it enables more scalable analysis through regression analysis (Lipman et al., 2023). For the estimated flow u(x, t),


스크린샷 2026-07-04 오후 1 16 15


⑵ If the regression analysis is perfect, then ρ = ρ’, so the continuity equation constraint is essentially no longer necessary.

⑶ Since ρ’(x, t) and u(x, t) are not given, the following equation is used:


스크린샷 2026-07-04 오후 1 16 56



Input: 2026.07.03 01:21

results matching ""

    No results matching ""