Timezone: »

Latent Diffusion Energy-Based Model for Interpretable Text Modelling
Peiyu Yu · Sirui Xie · Xiaojian Ma · Baoxiong Jia · Bo Pang · Ruiqi Gao · Yixin Zhu · Song-Chun Zhu · Ying Nian Wu

Latent space energy-based model has drawn growing interest in generative modelling in recent years. Fueled by its flexibility in formulation and strong modelling power of the latent space, recent works have been built upon it aiming at the interpretability of text modeling. However, latent space energy-based model also inherits some of the flaws of energy-based model: the degenerated MCMC sampling quality in practice can lead to not only poor generation quality but also instability in training, especially on data with complex latent structures. On another front, diffusion probabilistic models, inspired by non-equilibrium thermodynamics, have provided an appealing methodology that breaks down the heavy generation process and eases the burden of sampling. In this work, we show how the philosophy of diffusion probabilistic models becomes a cure for the vanilla latent space energy-based model, namely the latent diffusion energy-based model. Such a symbiosis between these classical models offers the best of both worlds: it learns a well-structured and meaningful latent space; it can be flexibly extended to scenarios where data labels are available as supervision signal; it demonstrates superior performance on interpretable text modelling over strong baselines.

Author Information

Peiyu Yu (University of California, Los Angeles)
Sirui Xie (UCLA)
Xiaojian Ma (Tsinghua University)

Xiaojian Ma is a senior undergraduate student at Tsinghua University and will be an incoming graduate student at the University of California, Los Angeles working with Prof. Song-chun Zhu. Previously he has been a visiting student at the National University of Singapore, with Prof. Gim Hee Lee. He also spends time at Tencent AI Lab, mentored by Dr. Wenbing Huang. His research interest including machine learning in general (statistical learning, learning theory, and reinforcement learning) and its application in robotics (imitation learning, optimal control, manipulation, grasping and multimodel sensing). Some of his research has been accepted as full papers or workshop papers to top-tier machine learning and robotics conferences as AAAI/ICRA/IROS/CVPR.

Baoxiong Jia (UCLA)
Bo Pang (University of California Los Angeles)
Ruiqi Gao (UCLA)
Yixin Zhu (Peking University)
Song-Chun Zhu (UCLA)
Ying Nian Wu (UCLA)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors