Continual Learning with Node-Importance based Adaptive Group Sparse Regularization

By Sangwon Jung, Hongjoon Ahn, Sungmin Cha, Taesup Moon. 2020

163 citations recorded by the source corpus.

We propose a novel regularization-based continual learning method, dubbed as Adaptive Group Sparsity based Continual Learning (AGS-CL), using two group sparsity-based penalties. Our method selectively employs the two penalties when learning each node based its the importance, which is adaptively updated after learning each new task. By utilizing the proximal gradient descent method for learning, the exact sparsity and freezing of the model is guaranteed, and thus, the learner can explicitly control the model capacity as the learning continues. Furthermore, as a critical detail, we re-initialize the weights associated with unimportant nodes after learning each task in order to prevent the negative transfer that causes the catastrophic forgetting and facilitate efficient learning of new tasks. Throughout the extensive experimental results, we show that our AGS-CL uses much less additional memory space for storing the regularization parameters, and it significantly outperforms several state-of-the-art baselines on representative continual learning benchmarks for both supervised and reinforcement learning tasks.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Optimization & Training, 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 Optimization & Training

In Meta & Continual Learning

Citations in the corpus

A Comprehensive Survey of Continual Learning: Theory, Method and Application (2023, background): “UCL [115] and AGS-CL [116] freeze the parameters connecting the important neurons, equivalent to a hard version of weight regularization.”
Incorporating neuro-inspired adaptability for continual learning in artificial intelligence (2023, positive): “AGS-CL [44] took advantages of parameter isolation and synaptic regularization to prevent catastrophic forgetting.”
CoSCL: Cooperation of Small Continual Learners is Stronger than a Big One (2022, positive): “Representative strategies include weight regularization [20,52,1], memory replay [36,44,46], parameter isolation [41,18] and dynamic architecture [14,50].”
NISPA: Neuro-Inspired Stability-Plasticity Adaptation for Continual Learning in Sparse Networks (2022, background): “There are several empirical results that support this choice, such as the observation that removing the most active units degrades performance much more than removing the same number of randomly chosen units (Jung et al., 2020a).”
Representation Compensation Networks for Continual Semantic Segmentation (2022, background): “Another research line [1, 44] explores the sparsity regularization for network parameters, which aims at activating as few neurons as possible for each task.”
AFEC: Active Forgetting of Negative Transfer in Continual Learning (2021, positive): “We reproduce the results of all the baselines from the ofﬁcially released code of [10], where we do an extensive hyperparameter search and report the best performance for fair comparison (detailed in Appendix C).”
CPR: Classifier-Projection Regularization for Continual Learning (2020, methodology): “We apply CPR to four regularization-based supervised CL methods: EWC (Kirkpatrick et al., 2017), SI (Zenke et al., 2017), MAS (Aljundi et al., 2018), RWalk (Chaudhry et al., 2018) and AGS-CL (Jung et al., 2020), and further analyze CPR via ablation studies and feature map visualizations.”
SS-IL: Separated Softmax for Incremental Learning (2020, methodology): “For general continual learning, there has been three main approaches; 1) regularizationbased [22, 34, 9, 2, 19], 2) dynamic architecture-based [29, 33, 24, 17], and 3) exemplar/replay-memory based [26, 10, 11, 30, 21, 32, 30, 21, 32] methods.”
Preserving Earlier Knowledge in Continual Learning with the Help of All Previous Feature Extractors (2021, background): “, kernels in CNNs) crucial for old knowledge, and then try to keep them unchanged with the help of regularization loss terms when learning new classes [5], [6], [7], [8], [9], [10], [11].”
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).”
Lifelong Event Detection with Knowledge Transfer (2021, methodology): “In addition to class incremental learning, regularization-based methods (Kirkpatrick et al., 2016; Jung et al., 2020; Ahn et al., 2019; Golkar et al., 2019; Serrà et al., 2018) are widely used for the situation where the model needs to learn a sequence of “disjoint” tasks.”
Understanding Catastrophic Forgetting and Remembering in Continual Learning with Optimal Relevance Mapping (2021, methodology): “The experimental implementation for most comparative methods mentioned in Table 3 have been taken from official implementations of [45], [50], [16] and [40].”
Formalizing the Generalization-Forgetting Trade-off in Continual Learning (2021, background): “For example, common solutions to the catastrophic forgetting issue include (1) representation-driven approaches [49, 25], (2) regularization approaches [27, 2, 34, 16, 48, 48, 24, 37, 10, 43], and (3) memory/experience replay [31, 32, 11, 12, 17].”
Lifelong Generative Modelling Using Dynamic Expansion Graph Model (2021, background): “The importance of processing modules during LLL was considered in (Aljundi, Kelchtermans, and Tuytelaars 2019; Jung et al. 2020).”
Effects of Auxiliary Knowledge on Continual Learning (2022, background): “Alternatively, structuring can be pursued at an architectural level, by explicitly allocating distinct subsets of model parameters to distinct tasks [8], [20], [21], encouraging nonoverlapping activation patterns for different data [9], [22], or by simply applying dropout [23], [24].”
Continuum: Simple Management of Complex Continual Learning Scenarios (2021, background): “) [22, 19] or tasks [20, 43], (incremental learning) either the current environment change while learning a single tasks (lifelong learning).”
Forget-free Continual Learning with Winning Subnetworks (2022, background): “…continual learning, which can be broadly categorized as follows: (1) Regularization-based methods (Kirkpatrick et al., 2017; Chaudhry et al., 2020; Jung et al., 2020; Tit-sias et al., 2020; Mirzadeh et al., 2021) aim to keep the learned information of past tasks during continual training aided by…”
Provable and Efficient Continual Representation Learning (2022, unknown): “…Buzzega et al. (2021); Aljundi et al. (2019); Borsos et al. (2020)), regularization-based (Kirkpatrick et al. (2017); Zenke et al. (2017); Li and Hoiem (2017); Jung et al. (2020)), and architecture-based methods (Fernando et al. (2017); Yoon et al. (2017); Mallya et al. (2018); Rusu et al. (2016)).”