Saliency Guided Experience Packing for Replay in Continual Learning

Saha, Gobinda; Roy, Kaushik

doi:10.1109/wacv56688.2023.00524

Cited by 8 publications

(4 citation statements)

References 30 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In this way, we could choose the smallest number of samples that contain the most influential information about each task. Several works in the literature propose methods for mitigating catastrophic forgetting in artificial neural networks that exploit explainable AI [32,33,34], and they manage to achieve comparable and, in some cases, higher performances than other state-of-the-art approaches.…”

Section: Discussionmentioning

confidence: 99%

Memory Replay For Continual Learning With Spiking Neural Networks

Proietti,

Ragno,

Capobianco

2023

2023 IEEE 33rd International Workshop on Machine Learning for Signal Processing (MLSP)

View full text Add to dashboard Cite

Two of the most impressive features of biological neural networks are their high energy efficiency and their ability to continuously adapt to varying inputs. On the contrary, the amount of power required to train top-performing deep learning models rises as they become more complex. This is the main reason for the increasing research interest in spiking neural networks, which mimic the functioning of the human brain achieving similar performances to artificial neural networks, but with much lower energy costs. However, even this type of network is not provided with the ability to incrementally learn new tasks, with the main obstacle being catastrophic forgetting. This paper investigates memory replay as a strategy to mitigate catastrophic forgetting in spiking neural networks. Experiments are conducted on the MNISTsplit dataset in both class-incremental learning and task-free continual learning scenarios.

show abstract

Section: Discussionmentioning

confidence: 99%

Memory Replay For Continual Learning With Spiking Neural Networks

Proietti,

Ragno,

Capobianco

2023

2023 IEEE 33rd International Workshop on Machine Learning for Signal Processing (MLSP)

View full text Add to dashboard Cite

show abstract

“…These methods retrieve samples from the memory buffer and combine them with the incoming data for model updates. Various strategies exist in this category, including experience replay (ER) [22], maximally interfered retrieval (MIR) [28], adversarial Shapley value experience replay (ASER) [23], gradient coreset replay (GCR) [24], and experience packing and replay (EPR) [25]. However, these methods primarily focus on data storage and may not fully consider the impact of new data, potentially leading to a decrease in classification performance.…”

Section: Methods Sampling Strategymentioning

confidence: 99%

“…There are three main continual learning approaches: regularization [14]- [16], parameter isolation [17]- [19], and replay-based learning methods [22]- [25]. Among them, replay-based learning methods are simple and effective and have been widely used in online class-incremental learning studies.…”

Section: Introductionmentioning

confidence: 99%

MixER: Mixup-Based Experience Replay for Online Class-Incremental Learning

Lim,

Zhou,

Kim

et al. 2024

IEEE Access

View full text Add to dashboard Cite

Continual learning in the online class-incremental setting aims to learn new classes continuously from a consistent data stream while retaining the knowledge of old classes to prevent catastrophic forgetting. Traditional replay-based methods store and use old-class data to achieve this. However, they often overlook the representation shift caused by the incoming data streams, which leads to suboptimal classification accuracy. In this study, we propose a solution for mitigating representation shifts by incorporating asymmetric mixup training into the replay method. Our approach is based on the concept that mixup-based training enhances the stability of model predictions and gradient norms between training samples. Our method differs from typical mixup augmentation, which is uniformly applied to all data. Instead, it selectively targets the old data stored in the memory buffer, deliberately excluding the classes from the newly incoming data. This approach enables the model to learn new data while preserving the representation of the old data. Moreover, our experiments demonstrate the effectiveness of the proposed method, which not only enhances the performance of replay-based methods but can also be seamlessly integrated as an additional compatible module into various replay-based techniques. Evaluation on the CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets demonstrates that our approach surpasses existing replay-based methods. It addresses the limitations of conventional replay techniques and offers a potential solution for continual learning scenarios. Our source code is publicly available at https://github.com/laymond1/MixER.

show abstract

“…One line of works (Kirkpatrick et al 2017;Schwarz et al 2018;Ebrahimi et al 2020;Saha et al 2021a;Kao et al 2021) achieves this goal by penalizing or preventing changes to the most important weights of the model for the past tasks while learning new tasks. Other works minimize forgetting by either storing samples from old tasks in the memory (Robins 1995;Lopez-Paz and Ranzato 2017;Chaudhry et al 2019bChaudhry et al , 2021Saha and Roy 2023) or synthesizing old data by generative models (Shin et al 2017) for rehearsal. Despite varying degrees of success, the stability-plasticity balance in such methods breaks down under long sequence of learning.…”

Section: Introductionmentioning

confidence: 99%

Continual Learning with Scaled Gradient Projection

Saha

Roy

2023

AAAI

View full text Add to dashboard Cite

In neural networks, continual learning results in gradient interference among sequential tasks, leading to catastrophic forgetting of old tasks while learning new ones. This issue is addressed in recent methods by storing the important gradient spaces for old tasks and updating the model orthogonally during new tasks. However, such restrictive orthogonal gradient updates hamper the learning capability of the new tasks resulting in sub-optimal performance. To improve new learning while minimizing forgetting, in this paper we propose a Scaled Gradient Projection (SGP) method, where we combine the orthogonal gradient projections with scaled gradient steps along the important gradient spaces for the past tasks. The degree of gradient scaling along these spaces depends on the importance of the bases spanning them. We propose an efficient method for computing and accumulating importance of these bases using the singular value decomposition of the input representations for each task. We conduct extensive experiments ranging from continual image classification to reinforcement learning tasks and report better performance with less training overhead than the state-of-the-art approaches.

show abstract

Saliency Guided Experience Packing for Replay in Continual Learning

Cited by 8 publications

References 30 publications

Memory Replay For Continual Learning With Spiking Neural Networks

Memory Replay For Continual Learning With Spiking Neural Networks

MixER: Mixup-Based Experience Replay for Online Class-Incremental Learning

Continual Learning with Scaled Gradient Projection

Contact Info

Product

Resources

About