Trans-Dimensional Generative Modeling via Jump Diffusion Models
Trans-Dimensional Generative Modeling via Jump Diffusion Models
复制标题
DOI：
10.48550/arxiv.2305.16261
复制
发表时间：
2023-05
期刊：
影响因子：
--
通讯作者：
Andrew Campbell;William Harvey;Christian Weilbach;Valentin De Bortoli;Tom Rainforth;A. Doucet
中科院分区：
文献类型：
--
作者：
Andrew Campbell;William Harvey;Christian Weilbach;Valentin De Bortoli;Tom Rainforth;A. Doucet
我们提出了一类新的生成模型，通过对每个数据点的状态和维度进行联合建模，自然地处理不同维度的数据。生成过程被公式化为在不同维度空间之间进行跳跃的跳跃扩散过程。我们首先定义了一个维度破坏正向噪声过程，然后推导出维度创建时间反向生成过程沿着一个新的证据下限训练目标，以学习近似它。模拟我们学习的时间反向生成过程的近似，然后提供了一种有效的方式，通过联合生成状态值和维度的采样数据的不同维度。我们在不同维度的分子和视频数据集上展示了我们的方法，报告了与测试时间扩散指导插补任务的更好兼容性，以及与分别生成状态值和维度的固定维度模型相比改进的插值功能。
We propose a new class of generative models that naturally handle data of varying dimensionality by jointly modeling the state and dimension of each datapoint. The generative process is formulated as a jump diffusion process that makes jumps between different dimensional spaces. We first define a dimension destroying forward noising process, before deriving the dimension creating time-reversed generative process along with a novel evidence lower bound training objective for learning to approximate it. Simulating our learned approximation to the time-reversed generative process then provides an effective way of sampling data of varying dimensionality by jointly generating state values and dimensions. We demonstrate our approach on molecular and video datasets of varying dimensionality, reporting better compatibility with test-time diffusion guidance imputation tasks and improved interpolation capabilities versus fixed dimensional models that generate state values and dimensions separately.