IMPLICIT STOCHASTIC PROCESSES, AND FUNCTION SPACE INFERENCE
Many supervised learning techniques in modern ML rely on i.i.d. assumptions of training data points gathered from the system of interest. On the contrary, Bayesian Nonparametric methods (BNs) generalized this assumption by defining priors over prediction functions (i.e., the stochastic processes), which induces exchangeable priors over data points. This introduces correlations between data points, and the prediction tasks can be viewed as the Bayesian inference problem in function space.
One famous example (and perhaps the mostly used BN approach in ML) is the Gaussian Processes (GPs). This approach has proven to be able to provide relatively accurate and reliable uncertainty estimates on unseen data, which is crucial for real-world applications. However, GPs suffers from limited scalability and expressiveness, which limit the applications of GPs. Inspired by recent advances in implicit models, I tried to take advantage of implicit models to construct stochastic processes that are more expressive to GPs. Similar to the construction of Gaussian processes (GPs), we can assign implicit distributions over any finite collections of random variables. Then, similar to Bayesian non-parameterics, one can directly perform Bayesian inference in function space. Such models are called implicit processes (IP). IPs brings together the best of both worlds, that is, combining the elegance of inference, and the well-calibrated uncertainty of Bayesian nonparametric methods, with the strong representational power of implicit models.
The main difficulty for applying this function-space view point to practical ML problems is that, IPs generally does not have simple analytic solutions for Bayesian inference. The paper titled “Variational Implicit Processes (VIP)” was our initial attempt at solving this problem. We derived a novel and efficient function space inference pipeline for such implicit processes, that gives a closed-form approximation to the IP posterior. For applications, we successfully applied our method to the problem of function space Bayesian inference for complicated IP priors. Click here for more information.
In the VIP paper, we discovered that the function space inference method applied to parameteric models (such as BNNs, Bayesian LSTMs, and dynamic systems) is very useful, and it often outperforms weight space VI methods. In line with the function-space perspective provided by VIP, we further investigate function space variational inference methods for general stochastic processes, that are more principled, accurate, and scalable.
Variational Implicit Processes Chao Ma, Yingzhen Li, and José Miguel Hernández-Lobato. Proceedings of the 36th International Conference on Machine Learning, 2019.
Functional Variational Inference for Stochastic Processes Chao Ma, José Miguel Hernández-Lobato. Submitted, 2020
DATA EFFICIENT MACHINE LEARNING
State of the art modern AI algorithms often rely on the power of large datasets. One the contrary, Bayesian methods are often considered to be able to handle small-data regime much better. How about the gap in-between these two extreme situations? For example, human experts are often able to: 1, possess exaustive domain knowledge, 2, dynamically acquire new information based on the current understanding of the situation, and 3, make effective decisions given such limited information. Imagine a person walking into a hospital with a broken arm. The first question from health-care personnel would likely be “How did you break you arm?” instead of “Do you have a cold?”, because the answer reveals relevant information for this patient’s treatment.
Automating this human expertise of asking relevant questions is the key towards data efficient machine learning and decision making. However, in many real-world applications it is extremely difficult. In this research, we use Bayesian deep learning methods and expermental design to build powerful models and achieve optimal trade-offs.
EDDI: Efficient Dynamic Discovery of High-value Information with Partial VAEChao Ma, Sebastian Tschiatschek, Konstantina Palla, José Miguel Hernández-Lobato, Sebastian Nowozin, and Cheng Zhang. Proceedings of the 36th International Conference on Machine Learning, 2019.
FIT: a Fast and Accurate Framework for Solving Medical Inquiring and Diagnosing Tasks Weijie He, Xiaohao Mao, Chao Ma, José Miguel Hernández-Lobato, and Ting Chen. submitted, 2020
Bayesian EDDI: Sequential Variable Selection with Bayesian Partial VAE Chao Ma*, Wenbo Gong*, Sebastian Tschiatschek, Sebastian Nowozin, José Miguel Hernández-Lobato, and Cheng Zhang. Real-world Sequential Decision Making workshop, International Conference on Machine Learning, 2019.
GENERATIVE MODELS FOR TABULAR DATA
Deep generative models often perform poorly in real-world applications due to the heterogeneity of natural data sets. Heterogeneity arises from data containing different types of features (categorical, ordinal, continuous, etc.) and features of the same type having different marginal distributions. We propose an extension of variational autoencoders (VAEs) called VAEM to handle such heterogeneous data. VAEM is a deep generative model that is trained in a two stage manner such that the first stage provides a more uniform representation of the data to the second stage, thereby sidestepping the problems caused by heterogeneous data. We provide extensions of VAEM to handle partially observed data, and demonstrate its performance in data generation, missing data prediction and sequential feature selection tasks. Our results show that VAEM broadens the range of real-world applications where deep generative models can be successfully deployed.
VAEM: a Deep Generative Model for Heterogeneous Mixed Type Data Chao Ma, Sebastian Tschiatschek, Richard Turner, José Miguel Hernández-Lobato, and Cheng Zhang. Advances in Neural Information Processing Systems, 2020.
HM-VAEs: a Deep Generative Model for Real-valued Data with Heterogeneous Marginals Chao Ma, Sebastian Tschiatschek, Yingzhen Li, Richard Turner, José Miguel Hernández-Lobato, and Cheng Zhang. Proceedings of The 2nd Symposium on Advances in Approximate Bayesian Inference, 2020.
Partial VAE for Hybrid Recommender System Chao Ma*, Wenbo Gong*, José Miguel Hernández-Lobato, Noam Koenigstein, Sebastian Nowozin, and Cheng Zhang. Neural Information Processing Systems Workshop on Bayesian Deep Learning, 2018.