Dynamic Latent Separation for Deep Learning
This work addresses the need for more flexible and interpretable latent representations in machine learning, offering a general approach that is not application-specific, though it appears incremental in nature.
The paper tackles the problem of learning expressive and interpretable latent variables for complex data by introducing a dynamic latent separation method that enhances output diversity and provides partial interpretation without supervision. The result is improved performance across classification and generation tasks for models of varying scales.
A core problem in machine learning is to learn expressive latent variables for model prediction on complex data that involves multiple sub-components in a flexible and interpretable fashion. Here, we develop an approach that improves expressiveness, provides partial interpretation, and is not restricted to specific applications. The key idea is to dynamically distance data samples in the latent space and thus enhance the output diversity. Our dynamic latent separation method, inspired by atomic physics, relies on the jointly learned structures of each data sample, which also reveal the importance of each sub-component for distinguishing data samples. This approach, atom modeling, requires no supervision of the latent space and allows us to learn extra partially interpretable representations besides the original goal of a model. We empirically demonstrate that the algorithm also enhances the performance of small to larger-scale models in various classification and generation problems.