Generative Feature Replay For Class-Incremental Learning

By Xialei Liu, Chenshen Wu, Mikel Menta, Luis Herranz, B. Raducanu, Andrew D. Bagdanov, Shangling Jui, Joost van de Weijer. 2020

190 citations recorded by the source corpus.

Humans are capable of learning new tasks without forgetting previous ones, while neural networks fail due to catastrophic forgetting between new and previously-learned tasks. We consider a class-incremental setting which means that the task-ID is unknown at inference time. The imbalance between old and new classes typically results in a bias of the network towards the newest ones. This imbalance problem can either be addressed by storing exemplars from previous tasks, or by using image replay methods. However, the latter can only be applied to toy datasets since image generation for complex datasets is a hard problem.We propose a solution to the imbalance problem based on generative feature replay which does not require any exemplars. To do this, we split the network into two parts: a feature extractor and a classifier. To prevent forgetting, we combine generative feature replay in the classifier with feature distillation in the feature extractor. Through feature generation, our method reduces the complexity of generative replay and prevents the imbalance problem. Our approach is computationally efficient and scalable to large datasets. Experiments confirm that our approach achieves state-of-the-art results on CIFAR-100 and ImageNet, while requiring only a fraction of the storage needed for exemplar-based continual learning. Code available at https://github.com/xialeiliu/GFR-IL.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Generative Models, Meta & Continual Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Generative Models

In Meta & Continual Learning

Citations in the corpus

A Comprehensive Survey of Continual Learning: Theory, Method and Application (2023, methodology): “For example, GFR [172] trains a conditional GANs to replay generated features after the feature extractor.”
Memory Replay with Data Compression for Continual Learning (2022, positive): “Another solution is feature replay: GFR (Liu et al., 2020a) learned a feature generator from a feature extractor to replay generated features, but the feature extractor suffered from catastrophic forgetting since it was incrementally updated.”
Model Zoo: A Growing Brain That Learns Continually (2021, methodology): “…1995; Shin et al., 2017; Lee et al., 2017), new architectures (Serra et al., 2018), generative replay-based (Mocanu et al., 2016; Shin et al., 2017; Liu et al., 2020; Ven et al., 2020), ensemble-based (Aljundi et al., 2017; Wen et al., 2020) and methods that select locally-redundant directions in…”
Training Networks in Null Space of Feature Covariance for Continual Learning (2021, methodology): “…GAN [56], MCIL [30], LwM [10], etc. Due to inaccessibility to the full datasets of previous tasks, they commonly use data of current task [10, 27], external data [24], coreset of previous tasks [30, 44] or synthetic data [56], resulting in distributional shift [28] to the original datasets.”
ORDisCo: Effective and Efficient Usage of Incremental Unlabeled Data for Semi-supervised Continual Learning (2021, background): “In the next task or batch, the generator replays conditional samples with the new training samples to update the generator and the classifier [35, 43, 45, 18].”
RATT: Recurrent Attention to Transient Tasks for Continual Image Captioning (2020, methodology): “Approaches include methods that mitigate catastrophic forgetting via replay of exemplars (iCarl [32], EEIL [3], and GEM [22]) or by performing pseudo-replay with GAN-generated data [21, 39, 43].”
Class-Incremental Learning: Survey and Performance Evaluation on Image Classification (2020, background): “To address this problem, some recent works have proposed to perform feature replay instead of image replay [49], [100], [101], where instead a generator is trained to generate features at some hidden layer of the network.”
Memory-Efficient Semi-Supervised Continual Learning: The World is its Own Replay Buffer (2021, neutral): “Rather than storing raw data, some methods train a generative model [40, 41, 42] or replay compressed data representations in a late layer [43].”
Technical Report for ICCV 2021 Challenge SSLAD-Track3B: Transformers Are Better Continual Learners (2022, methodology): “To overcome catastrophic forgetting in artificial neural networks, three types of methods have been explored in the past few years, replay [15, 3, 6, 20, 10], regularization [9, 19, 8, 4, 17, 13, 5, 7, 16], and expand [22, 14, 1, 21].”
Long Live the Lottery: The Existence of Winning Tickets in Lifelong Learning (2021, background): “And some generative lifelong learning methods (Liu et al., 2020; Shin et al., 2017) mitigate catastrophic forgetting by generating simulated data of previous tasks.”
Continual Detection Transformer for Incremental Object Detection (2023, methodology): “ER methods [6, 30, 34, 35, 43, 54] work instead by memorising some of the past training data (the exemplars ), replaying them in the following phases to “re-member” the old object categories.”
Towards Exemplar-Free Continual Learning in Vision Transformers: an Account of Attention, Functional and Weight Regularization (2022, background): “When the regularization takes place in intermediate layers, the model can experience excessive stability, therefore loosing in plasticity abilities [14,24,51].”
Generative Feature Replay with Orthogonal Weight Modification for Continual Learning (2020, unknown): “4) For GFR, we include naive GFR which directly uses GFR without any other CL method and FD+GFR (Liu et al. 2020).”
Dual-Teacher Class-Incremental Learning With Data-Free Generative Replay (2021, methodology): “This is different from conventional generative replay methods [19, 29, 32], where auxiliary generative models are trained along with the main objective models by using the original training data, and the pre-trained generative models are shared for later incremental learning.”
Class-Incremental Learning with Generative Classifiers (2021, background): “An advantage of this second approach, which is reminiscent of recent studies that performed generative replay in the feature space [30, 50], is that it appears to be easier to learn good generative models for such extracted features, presumably because they are less complex than the raw inputs.”