Elizabeth L. Baker,Moritz Schauer,Stefan Sommer
Abstract:We propose a new algorithm for learning a bridged diffusion process using score-matching methods. Our method relies on reversing the dynamics of the forward process and using this to learn a score function, which, via Doob's $h$-transform, gives us a bridged diffusion process; that is, a process conditioned on an endpoint. In contrast to prior methods, ours learns the score term $\nabla_x \log p(t, x; T, y)$, for given $t, Y$ directly, completely avoiding the need for first learning a time reversal. We compare the performance of our algorithm with existing methods and see that it outperforms using the (learned) time-reversals to learn the score term. The code can be found at <a class="link-external link-https" href="https://github.com/libbylbaker/forward_bridge" rel="external noopener nofollow">this https URL</a>.
What problem does this paper attempt to address?
### What problem does this paper attempt to solve?
This paper aims to solve the problem of how to learn and simulate the conditional diffusion process (i.e., the bridging diffusion process). Specifically, the author proposes a new algorithm that uses the score - matching method to learn the score term \(\nabla_x \log p(t, x; T, y)\) in the bridging diffusion process, given time \(T\) and end - point \(y\).
#### Main problem description:
1. **Simulation of the conditional diffusion process**:
- Given a diffusion process, we hope that at a specific time \(T\), the process can reach a specific value or distribution. This can be represented as a new stochastic differential equation (SDE) through Doob's \(h\)-transformation of the conditional diffusion process, but this new SDE contains a difficult - to - handle score term \(\nabla_x \log p(t, x; T, y)\).
2. **Learning of the score term**:
- Traditional methods usually need to first learn the time - reversed diffusion process and then use it to learn the score term. This method introduces additional errors and has a high computational complexity.
- The method proposed in this paper directly learns the score term \(\nabla_x \log p(t, x; T, y)\), completely avoiding the learning of time - reversal, thereby reducing errors and improving efficiency.
3. **Application background**:
- Diffusion processes are widely used in multiple fields, such as medical imaging, evolutionary biology, etc. In particular, in shape analysis, the bridging process is used to model the process of organ shape changes with diseases, or the shape changes of organisms over time.
### Method innovation points:
- **Avoiding time - reversal**: By combining the score - matching method and the adjoint diffusion process, directly learn the score term without first learning the time - reversed diffusion process.
- **Reducing errors**: Compared with existing methods, this method only requires one training, reducing the cumulative errors caused by two trainings.
- **Flexibility**: It can be extended to multiple end - point conditions, including fixed end - points and end - point distributions.
### Experimental verification:
The author verifies the effectiveness of this method through a series of experiments, including the Ornstein - Uhlenbeck process, Brownian motion, and cell differentiation models. The experimental results show that this method has a lower error when simulating the forward bridging diffusion process and is superior to the existing time - reversal methods.
### Summary:
This paper proposes a new algorithm that uses the score - matching method to directly learn the score term in the bridging diffusion process, solving the complexity and error problems caused by the need for time - reversal in traditional methods. This method performs well in multiple application scenarios and has important theoretical and practical significance.