CL-BioGAN: Biologically-Inspired Cross-Domain Continual Learning for Hyperspectral Anomaly Detection

TL;DR

This work tackles the challenge of forgetting in open-domain cross-scene Hyperspectral Anomaly Detection by introducing CL-BioGAN, a biologically inspired continual learning GAN. It combines a BioGAN with a self-attention discriminator, a bio-inspired CL loss, and a replay strategy to balance memory stability and learning plasticity across sequential HAD tasks. The approach leverages a Spectral-Based Background Selection Strategy, Spatial-Spectral Noise Smoothing, and a Bayesian-inspired forgetting mechanism (AF loss) to improve robust background reconstruction and anomaly detection with fewer parameters. Experiments on ABU and HAD100 datasets demonstrate improved ACC and reduced forgetting, illustrating the practical impact of biologically motivated forgetting and replay in open-world hyperspectral analysis.

Abstract

Memory stability and learning flexibility in continual learning (CL) is a core challenge for cross-scene Hyperspectral Anomaly Detection (HAD) task. Biological neural networks can actively forget history knowledge that conflicts with the learning of new experiences by regulating learning-triggered synaptic expansion and synaptic convergence. Inspired by this phenomenon, we propose a novel Biologically-Inspired Continual Learning Generative Adversarial Network (CL-BioGAN) for augmenting continuous distribution fitting ability for cross-domain HAD task, where Continual Learning Bio-inspired Loss (CL-Bio Loss) and self-attention Generative Adversarial Network (BioGAN) are incorporated to realize forgetting history knowledge as well as involving replay strategy in the proposed BioGAN. Specifically, a novel Bio-Inspired Loss composed with an Active Forgetting Loss (AF Loss) and a CL loss is designed to realize parameters releasing and enhancing between new task and history tasks from a Bayesian perspective. Meanwhile, BioGAN loss with L2-Norm enhances self-attention (SA) to further balance the stability and flexibility for better fitting background distribution for open scenario HAD (OHAD) tasks. Experiment results underscore that the proposed CL-BioGAN can achieve more robust and satisfying accuracy for cross-domain HAD with fewer parameters and computation cost. This dual contribution not only elevates CL performance but also offers new insights into neural adaptation mechanisms in OHAD task.

Figures (8)

• Figure 1: Comparison of traditional DL training model and the CL method. (a) represents the traditional DL method, which obtains OHAD results by a set of independent well-trained parameters. (b) represents the proposed CL method. The parameters of the model are continuously updated with the arrived tasks, but the updated parameters will not forget the previous learned knowledge.
• Figure 2: Principles of active forgetting and learning in our proposed neural network. (a) represents the Neural and Protrusion Function Diagram. (b) represents the Synaptic Changes Diagram During the Learning of a New Task.
• Figure 3: Overview of the proposed CL-BioGAN for OHAD. The left part represents the continual training phase, which mainly composed with Replay Strategy, BioGAN structure and Bio-Inspired Loss (CL Loss, AF Loss and recon Loss). The right part in the figure is continual testing part, the well-trained Bio-Generator is only used in testing phase to detect the continually arrived HAD scenarios.
• Figure 4: The catastrophic forgetting performance evaluation for open scenario OHAD on previous tasks. (a) AUC value for tasks 1-2 after training on the second task. (b) AUC value for tasks 1-3 after training on the third task. (c) AUC value for tasks 1-4 after training on the fourth task. (d) AUC value for tasks 1-5 after training on the fifth task.
• Figure 5: The ACC, BWT and FWT Performance evaluation for OHAD of different methods. (a) The variation of $ACC$ for Task 1-5. (b) The variation of $BWT$ for OHAD Task 1-5. (c) The variation of $FWT$ for OHAD Task 1-5.