Abstract:Conditional generative models represent a significant advancement in the field of machine learning, allowing for the controlled synthesis of data by incorporating additional information into the generation process. In this work we introduce a novel Schrödinger bridge based deep generative method for learning conditional distributions. We start from a unit-time diffusion process governed by a stochastic differential equation (SDE) that transforms a fixed point at time $0$ into a desired target conditional distribution at time $1$. For effective implementation, we discretize the SDE with Euler-Maruyama method where we estimate the drift term nonparametrically using a deep neural network. We apply our method to both low-dimensional and high-dimensional conditional generation problems. The numerical studies demonstrate that though our method does not directly provide the conditional density estimation, the samples generated by this method exhibit higher quality compared to those obtained by several existing methods. Moreover, the generated samples can be effectively utilized to estimate the conditional density and related statistical quantities, such as conditional mean and conditional standard deviation.
What problem does this paper attempt to address?
### Problems the paper attempts to solve
The paper aims to solve the problem of high - quality sample generation in conditional generation models. Specifically, the paper proposes a deep conditional generation method based on Schrödinger Bridge (SB) for learning conditional distributions. Traditional methods face the problems of high computational complexity and low sample quality when generating high - dimensional conditional samples. By introducing SB technology, this paper improves the performance of the conditional generation model, making the generated samples of higher quality and can be effectively used to estimate conditional densities and related statistics.
### Main contributions
1. **Propose a new generation method**: The paper proposes a deep generation method based on SB, which transforms the initial point at time 0 to the target conditional distribution at time 1 through a unit - time diffusion process.
2. **Non - parametric estimation**: Use deep neural networks to non - parametrically estimate the drift term in the diffusion process, and effectively implement it by discretizing the SDE through the Euler - Maruyama method.
3. **Wide application**: This method is applicable not only to low - dimensional conditional generation problems, but also to high - dimensional conditional generation problems.
4. **Superior performance**: Numerical studies show that although this method does not directly provide conditional density estimation, the quality of the generated samples is higher than that of existing methods, and can be effectively used to estimate conditional densities and related statistics, such as conditional means and conditional standard deviations.
### Method overview
1. **Schrödinger Bridge problem**: The SB problem seeks the most likely evolution path from the initial distribution to the target distribution within a finite time. The paper first reviews the SB problem and its application in unconditional sample generation.
2. **Conditional generation model**: By combining SB technology with the conditional generation model, the paper proposes a new framework for generating conditional samples. This framework estimates the drift term by optimizing the objective function, thereby generating high - quality conditional samples.
3. **Implementation details**: The paper describes in detail the training and generation processes, including using deep neural networks to estimate the drift term and generating samples by discretizing the SDE through the Euler - Maruyama method.
### Numerical experiments
1. **Two - dimensional sample visualization**: The effectiveness of the method is verified by generating three non - linear, non - Gaussian two - dimensional sample sets. The results show that the samples generated by the SBCG method are close to the target distribution.
2. **Complex data sets**: Tests were carried out on complex two - dimensional data sets such as checkerboard, moons, pinwheel and swissroll, and the results show that the samples generated by SBCG are highly consistent with the real distribution.
3. **Conditional mean and standard deviation estimation**: Compared with existing conditional density estimation methods (such as GCDS, CKDE, NNKCDE and FlexCode), SBCG performs excellently in estimating conditional means and conditional standard deviations.
### Conclusion
The paper proposes a deep conditional generation method based on SB, which effectively solves the high - dimensional conditional generation problem, generates high - quality samples, and can be used to estimate conditional densities and related statistics. The performance of this method on multiple data sets is better than that of existing methods and has broad application prospects.