CPR: Classifier-Projection Regularization for Continual Learning

By Sungmin Cha, Hsiang Hsu, Flávio du Pin Calmon, Taesup Moon. 2020

91 citations recorded by the source corpus.

We propose a general, yet simple patch that can be applied to existing regularization-based continual learning methods called classifier-projection regularization (CPR). Inspired by both recent results on neural networks with wide local minima and information theory, CPR adds an additional regularization term that maximizes the entropy of a classifier's output probability. We demonstrate that this additional term can be interpreted as a projection of the conditional probability given by a classifier's output to the uniform distribution. By applying the Pythagorean theorem for KL divergence, we then prove that this projection may (in theory) improve the performance of continual learning methods. In our extensive experimental results, we apply CPR to several state-of-the-art regularization-based continual learning methods and benchmark performance on popular image recognition datasets. Our results demonstrate that CPR indeed promotes a wide local minima and significantly improves both accuracy and plasticity while simultaneously mitigating the catastrophic forgetting of baseline continual learning methods.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, 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

1. Jun Zhu
2. Liyuan Wang
3. Xingxing Zhang
4. Hang Su
5. Jorge Armando Mendez Mendez
6. Eric Eaton
7. Liu Liu
8. Wei He

In Computer Vision and Pattern Recognition

1. Jun Zhu
2. Liyuan Wang
3. Xingxing Zhang
4. Hang Su
5. Dacheng Tao
6. P. Heng
7. Liu Liu
8. Wei He

In Meta & Continual Learning

1. Liyuan Wang
2. Xingxing Zhang
3. Jun Zhu
4. Hang Su
5. Jorge Armando Mendez Mendez
6. Eric Eaton
7. Yi Zhong
8. Qian Li

Citations in the corpus

• A Comprehensive Survey of Continual Learning: Theory, Method and Application (2023, background): “A distinctive feature of CRL is its high diversity of applications and benchmarks, including continuous control [203], [279], [296], [449], maze navigation [274], [370], video games (Atari [10], [59], [195], [220], [370], [429], StarCraft [372], Minecraft [405], Hanabi [312]), etc.”
• Incorporating neuro-inspired adaptability for continual learning in artificial intelligence (2023, methodology): “CAF (ours) and CPR [46] can be plug-and-play for regularization-based methods such as EWC [7] and MAS [8].”
• Balancing Stability and Plasticity through Advanced Null Space in Continual Learning (2022, methodology): “…well-established methods 4 , including 5 regularization-based methods using importance measure (EWC [17], MAS [1], MUC-MAS [25], SI [58], and CPR [5]), 2 regularization-based methods using knowledge distillation (LwF [23] and GD-WILD [20]), 1 architecture-based method (InstAParam [9]), and 6…”
• CoSCL: Cooperation of Small Continual Learners is Stronger than a Big One (2022, neutral): “Implementation: We mainly focus on the task-incremental setting used in [45,18,6,41] and follow their implementation for most experiments if not specified.”
• GCR: Gradient Coreset based Replay Buffer Selection for Continual Learning (2021, neutral): “CL approaches divide broadly into three main themes: Regularization based approaches preserve learning from earlier tasks via regularization terms placed on model weights (EWC [12], synaptic intelligence [13]), structural regularization [14], or functional regularization [15, 4]).”
• Flattening Sharpness for Dynamic Gradient Projection Memory Benefits Continual Learning (2021, methodology): “This involves using knowledge distillation [23, 14] or penalizing changes in weights deemed important for previous tasks [17, 46, 28, 2, 3] to reduce forgetting.”
• AlterSGD: Finding Flat Minima for Continual Learning by Alternative Training (2021, background): “Actually, there are some concurrent works also associating the flat local minimal with forgetting mitigation in CL from different viewpoints (e.g., in CPR [3] and stable SGD [31]).”
• TARGET: Federated Class-Continual Learning via Exemplar-Free Distillation (2023, methodology): “Inspired by Backward Transfer (BwT) [6], we derive the following formula to measure the severity of forgetting, a popular forgetting measure in CL [4, 7, 8]:”
• DDGR: Continual Learning with Deep Diffusion-based Generative Replay (2023, methodology): “There are also other regularization-based methods that are based on functional regularization (Pan et al., 2020; Benjamin et al., 2019), uncertainty regularization (Ahn et al., 2019), classifier projection regularization (Cha et al., 2021) and node importance (Jung et al., 2020).”
• A Unified and General Framework for Continual Learning (2024, methodology): “(Kirkpatrick et al., 2017; Zenke et al., 2017b; Chaudhry et al., 2018; Aljundi et al., 2018; Cha et al., 2021; Wang et al., 2021; Yang et al., 2023a).”
• How to Reuse and Compose Knowledge for a Lifetime of Tasks: A Survey on Continual Learning and Functional Composition (2022, methodology): “The downside is that the agent does not benefit from prior experience during initial exploration, which is critical for data-efficient training in lifelong RL.”