Continual Variational Autoencoder Learning via Online Cooperative Memorization

Abstract

Due to their inference, data representation and reconstruction properties, Variational Autoencoders (VAE) have been successfully used in continual learning classification tasks. However, their ability to generate images with specifications corresponding to the classes and databases learned during Continual Learning (CL) is not well understood and catastrophic forgetting remains a significant challenge. In this paper, we firstly analyze the forgetting behaviour of VAEs by developing a new theoretical framework that formulates CL as a dynamic optimal transport problem. This framework proves approximate bounds to the data likelihood without requiring the task information and explains how the prior knowledge is lost during the training process. We then propose a novel memory buffering approach, namely the Online Cooperative Memorization (OCM) framework, which consists of a Short-Term Memory (STM) that continually stores recent samples to provide future information for the model, and a Long-Term Memory (LTM) aiming to preserve a wide diversity of samples. The proposed OCM transfers certain samples from STM to LTM according to the information diversity selection criterion without requiring any supervised signals. The OCM framework is then combined with a dynamic VAE expansion mixture network for further enhancing its performance.

Task-Free Dynamic Sparse Vision Transformer for Continual Learning

“…However, most existing DEMs require knowing the task identity to provide the auxiliary information for the expansion strategy (Ye and Bors 2021). Recently, DEMs have been shown to achieve promising results in TFCL (Ye and Bors 2022a). The first study of using DEMs for TFCL was proposed in (Rao et al 2019), which introduces a continual learning framework called the Continual Unsupervised Representation Learning (CURL).…”

“…A similar idea called the Continual Neural Dirichlet Process Mixture (CN-DPM) uses the Dirichlet processes for the VAE component expansion (Lee et al 2020). Moreover DEMs can further improve their performance by using an efficient sample selection approach called the Online Cooperative Memorization (OCM) (Ye and Bors 2022a). which employs a a dual memory system to store both short and long-term information.…”

“…This drawback prompts several researchers to develop Dynamic Extension Models (DEMs) Bors 2021, 2022a), which increase their capacity by adding new models to cope with new tasks. One of the crucial advantages of DEMs over static models is that they can naturally solve the stability-plasticity dilemma by freezing previously learnt parameters and creating new components to learn new data distributions (Ye and Bors 2022a).…”

Task-Free Continual Generation and Representation Learning via Dynamic Expansionable Memory Cluster

“…Regularization and knowledge distillation methods have recently been considered to further improve performance (Aljundi et al 2019b;Chaudhry et al 2019b,a;Bang et al 2021Bang et al , 2022Cha, Lee, and Shin 2021;Kirkpatrick et al 2017;Kurle et al 2020;Li and Hoiem 2017). Although the memory-based approaches perform well, they suffer from the negative backward transfer when learning new samples (Ye and Bors 2022a). Such a drawback is solved by employing a dynamic network architecture which preserves prior knowledge into frozen parameters of specific units while building new hidden layers and units to learn novel tasks (Wen, Tran, and Ba 2020;Ye and Bors 2023b,a).…”

“…However, these methods perform worse when applied to learning a long-term data stream due to their fixed model capacity. Meanwhile, dynamic architecture models have achieved promising results (Lee et al 2020;Rao et al 2019;Ye and Bors 2022a). The first dynamic expansion model for TFCL, proposed in (Rao et al 2019), dynamically adds new inference models to capture data changes while using Generative Replay Mechanisms (GRMs) to alleviate forgetting.…”