Search research articles

ABOUT JoVE

Overview Leadership Blog JoVE Help Center

AUTHORS

Publishing Process Editorial Board Scope & Policies Peer Review FAQ Submit

LIBRARIANS

Testimonials Subscriptions Access Resources Library Advisory Board FAQ

RESEARCH

JoVE Journal Methods Collections JoVE Encyclopedia of Experiments Archive

EDUCATION

JoVE Core JoVE Business JoVE Science Education JoVE Lab Manual Faculty Resource Center Faculty Site

Terms & Conditions of Use

Search research articles

Related Experiment Videos

L³Net: Localized and Layered Reparameterization for incremental learning.

Xuandi Luo¹, Huaidong Zhang¹, Yi Xie¹

¹School of Future Technology, South China University of Technology, Guangzhou, China.

Neural Networks : the Official Journal of the International Neural Network Society

|March 28, 2025

Summary

This summary is machine-generated.

Related Concept Videos

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by

Same author

[Research progress of the CRISPR-Cas system in the detection and elimination of antibiotic resistance genes].

Sheng wu gong cheng xue bao = Chinese journal of biotechnology·2026

Same author

AvatarVTON: 4D Virtual Try-On for Animatable Avatars.

IEEE transactions on visualization and computer graphics·2026

Same author

Countering postoperative immune suppression with a self-assembling dendritic cell nanovaccine.

Bioactive materials·2026

Same author

How Far Are We from Generating Missing Modalities with Foundation Models?

IEEE transactions on pattern analysis and machine intelligence·2026

Same author

From Target-Nontarget to Nontarget Screening: A Review on Screening Methods for Organic Pollutants Based on High-Resolution Mass Spectrometry.

Environment & health (Washington, D.C.)·2026

Same author

Plasma proteomic profiles reveal proteins and characteristic patterns associated with hypertension: A prospective cohort study.

Journal of hypertension·2026

Same journal

Exploiting audio-visual modalities in videos: Object detection via multi-stage bilateral coupling network.

Neural networks : the official journal of the International Neural Network Society·2026

Same journal

Reliability-aware modality completion with cross-modal distillation for federated learning with missing modalities.

Neural networks : the official journal of the International Neural Network Society·2026

Same journal

IGFD-Net: Illumination-guided frequency decoupling for polarization image fusion.

Neural networks : the official journal of the International Neural Network Society·2026

Same journal

Multiple-Strategies dung beetle optimizer and its applications in engineering optimization and bankruptcy prediction.

Neural networks : the official journal of the International Neural Network Society·2026

Same journal

Aggregating global-scale pixel-wise forgery cues within a graph.

Neural networks : the official journal of the International Neural Network Society·2026

Same journal

Finite-Time intermittent control for secure synchronization of Neutral-Type stochastic delayed neural networks under aperiodic DoS attacks.

Neural networks : the official journal of the International Neural Network Society·2026

See all related articles

We introduce L³Net, a novel class incremental learning (CIL) framework that balances model complexity and performance. L³Net effectively mitigates catastrophic forgetting and class imbalance, outperforming existing CIL methods.

Area of Science:

Artificial Intelligence
Machine Learning
Computer Vision

Background:

Class incremental learning (CIL) methods combat catastrophic forgetting by parameter retention and architectural expansion.
Overly complex models and performance degradation during parameter compression pose significant challenges in CIL.

Purpose of the Study:

To propose a novel three-stage CIL framework, L³Net, balancing model complexity and performance.
To address catastrophic forgetting, inference overhead, and classification bias in CIL.

Main Methods:

Localized Dual-path Expansion: Integrates a fusion selector after each convolutional layer for simultaneous learning from old and new features.
Feature Selectors Gradient Resetting: Sparsifies fusion selectors to minimize conflicts between old and new features.

Keywords:

Class incremental learning Knowledge distillation Reparameterization

Related Experiment Videos

Decoupled Balanced Distillation and Logit Adjustment: Mitigates class imbalance and enhances knowledge retention from rehearsal data.

Main Results:

L³Net demonstrates superior performance compared to state-of-the-art methods on CIFAR-100 and ImageNet benchmarks.
The framework effectively balances model complexity and learning performance in incremental settings.

Conclusions:

L³Net offers an effective solution for class incremental learning, addressing key challenges of catastrophic forgetting and class imbalance.
The proposed methods provide a robust approach for building efficient and high-performing incremental learning systems.