Scholarlink

<p>Catastrophic forgetting was a phenomenon in artificial neural networks where a model rapidly and severely loses its performance on previous learned tasks. It happens because the network’s weights are updated to meet the new task’s objectives, which causes loss of information in previous tasks. Autonomous driving systems rely on the integration of high-performance computing, deep learning algorithms, and advanced sensors like LiDAR, cameras, and radar to process real-time environmental data and make instantaneous driving decisions. Various research works are conducted to develop autonomous vehicle image classification systems, but still numerous challenges were rectified. To address these challenges an innovative model is implemented in current autonomous vehicle classification. Initially, necessary images were acquired from standard dataset. These images were subsequently enhanced using Vision Transformer-Retinex (ViT-Retinex) to effectively manage diverse weather and light conditions. The pre-processed images undergo object detection using the Recurrent MobileNet Single Shot MultiBox Detector (RM-SSD), which is designed for consuming limited computational power. After detecting the objects, Incremental learning using the Adaptive Pyramid Dilated Mobilenetv3 Classifier (IL-APDMV3) is implemented to classify objects and mitigate catastrophic forgetting where new data acquisition compromises the retention of previous knowledge. The weights of the IL-APDMV3 model are optimally selected by Enhanced Random Variable-based Fossa Optimization Algorithm (ERV-FOA) for fine-tuning system’s superior object recognition and classification performance. Finally, performance of suggested method is validated against prior methods to prove effectiveness of system. This developed work has achieved 98.16% accuracy rate in K-fold 1. The framework demonstrates strong performance and superior results compared to recent techniques across several key evaluation metrics such as 10.66% FNR, 99.26% specificity, 93.06% F1-score, and 0.74% FPR.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Mitigating catastrophic forgetting in autonomous driving via an optimized incremental learning framework

  • 期刊论文

  • 发表日期：

    2026年5月11日

查看全文
  • Manimaran Mannaperumal,
  • Dhilipkumar Venkatesan

摘要

Catastrophic forgetting was a phenomenon in artificial neural networks where a model rapidly and severely loses its performance on previous learned tasks. It happens because the network’s weights are updated to meet the new task’s objectives, which causes loss of information in previous tasks. Autonomous driving systems rely on the integration of high-performance computing, deep learning algorithms, and advanced sensors like LiDAR, cameras, and radar to process real-time environmental data and make instantaneous driving decisions. Various research works are conducted to develop autonomous vehicle image classification systems, but still numerous challenges were rectified. To address these challenges an innovative model is implemented in current autonomous vehicle classification. Initially, necessary images were acquired from standard dataset. These images were subsequently enhanced using Vision Transformer-Retinex (ViT-Retinex) to effectively manage diverse weather and light conditions. The pre-processed images undergo object detection using the Recurrent MobileNet Single Shot MultiBox Detector (RM-SSD), which is designed for consuming limited computational power. After detecting the objects, Incremental learning using the Adaptive Pyramid Dilated Mobilenetv3 Classifier (IL-APDMV3) is implemented to classify objects and mitigate catastrophic forgetting where new data acquisition compromises the retention of previous knowledge. The weights of the IL-APDMV3 model are optimally selected by Enhanced Random Variable-based Fossa Optimization Algorithm (ERV-FOA) for fine-tuning system’s superior object recognition and classification performance. Finally, performance of suggested method is validated against prior methods to prove effectiveness of system. This developed work has achieved 98.16% accuracy rate in K-fold 1. The framework demonstrates strong performance and superior results compared to recent techniques across several key evaluation metrics such as 10.66% FNR, 99.26% specificity, 93.06% F1-score, and 0.74% FPR.