Timezone: »

BooVAE: Boosting Approach for Continual Learning of VAE
Evgenii Egorov · Anna Kuzina · Evgeny Burnaev

Wed Dec 08 12:30 AM -- 02:00 AM (PST) @

Variational autoencoder (VAE) is a deep generative model for unsupervised learning, allowing to encode observations into the meaningful latent space. VAE is prone to catastrophic forgetting when tasks arrive sequentially, and only the data for the current one is available. We address this problem of continual learning for VAEs. It is known that the choice of the prior distribution over the latent space is crucial for VAE in the non-continual setting. We argue that it can also be helpful to avoid catastrophic forgetting. We learn the approximation of the aggregated posterior as a prior for each task. This approximation is parametrised as an additive mixture of distributions induced by an encoder evaluated at trainable pseudo-inputs. We use a greedy boosting-like approach with entropy regularisation to learn the components. This method encourages components diversity, which is essential as we aim at memorising the current task with the fewest components possible. Based on the learnable prior, we introduce an end-to-end approach for continual learning of VAEs and provide empirical studies on commonly used benchmarks (MNIST, Fashion MNIST, NotMNIST) and CelebA datasets. For each dataset, the proposed method avoids catastrophic forgetting in a fully automatic way.

Author Information

Evgenii Egorov (University of Amsterdam)
Anna Kuzina (VU Amsterdam)
Evgeny Burnaev (Skoltech)

Evgeny is an experienced scientist working at the interface between machine learning and applied engineering problems. He obtained his Master’s degree in Applied Physics and Mathematics from the Moscow Institute of Physics and Technology in 2006. After successfully defending his PhD thesis in Foundations of Computer Science at the Institute for Information Transmission Problem RAS (IITP RAS) in 2008, Evgeny stayed with the Institute as a head of IITP Data Analysis and Modeling group. Today, Evgeny’s research interests encompass the areas of regression based on Gaussian Processes, bootstrap, confidence sets and conformal predictors, volatility modeling and nonparametric estimation, statistical decisions and rapid detection of anomalies in complex multicomponent systems. Evgeny always demonstrated a deep fundamental knowledge and engineer-like thinking that enabled him to effectively use methods of statistics, machine learning and predictive modeling to deal with practical tasks in hi-tech industries, primarily aerospace, medicine and life sciences. He carried out a number of successful industrial projects with Airbus, Eurocopter and Sahara Force India Formula 1 team among others. The corresponding data analysis algorithms, developed by Evgeny and his group at IITP, formed a core of the algorithmic software library for surrogate modeling and optimization. Thanks to the developed functionality, engineers can construct fast mathematical approximations to long running computer codes (realizing physical models) based on available data and perform design space exploration for trade-off studies. The software library passed the final Technology Readiness Level 6 certification in Airbus. According to Airbus experts, application of the library “provides the reduction of up to 10% of lead time and cost in several areas of the aircraft design process”. Nowadays several dozens of Airbus departments use it. Later a spin-off company developed a Software platform for Design Space Exploration with GUI based on this algorithmic core. Evgeny has also a considerable teaching experience both in Russian and English. He has developed and taught various undergraduate and graduate courses in applied mathematics at MIPT, IITP, Yandex School of Data Analysis and the Humboldt University of Berlin, as well as mini courses on application of machine learning in engineering multidisciplinary modeling and optimization for technological companies such as Astrium, Safran, SAFT, CNES, etc. Before joining Skoltech, Evgeny was a Lecturer at Yandex School of Data Analysis, Associate Professor and Vice Chairman of Information Transmission Problems and Data Analysis Chair at MIPT, data analysis expert at DATADVANCE llc., and head of IITP Data Analysis and Predictive Modeling Lab. At Skoltech, Evgeny is actively engaged in the development of CDISE educational and research programs, and continues his research in the areas of development of theoretical tools for estimation of change-point algorithms’ performance, effective algorithms for anomaly detection and failures prediction, analysis of their properties, and development of a core library for anomaly detection and failures prediction.

More from the Same Authors