受人脑中记忆机制启发的增量目标检测方法  

Incremental Object Detection Inspired by Memory Mechanisms in Brain

作　　者：商迪 吕彦锋 乔红[1,2,3] SHANG Di;LYU Yanfeng;QIAO Hong(State Key Laboratory of Multimodal Artificial Intelligence System,Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China;School of Artificial Intelligence,University of Chinese Academy of Sciences,Beijing 100049,China;State Key Laboratory of Complex System Management and Control,Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]中国科学院自动化研究所多模态人工智能系统全国重点实验室,北京100190 [2]中国科学院大学人工智能学院,北京100049 [3]中国科学院自动化研究所复杂系统管理与控制国家重点实验室,北京100190

出　　处：《计算机科学》2023年第2期267-274,共8页Computer Science

基　　金：北京市自然科学基金(L211023);科技创新2030-“新一代人工智能”重大项目(2020AAA0105900);国家自然科学基金(91948303)。

摘　　要：增量学习是缩小当前人工智能和人类智能间差距的关键技术,指智能体像人类一样从不稳定数据流中顺序学习多个任务,且不发生遗忘。目标检测是计算机视觉领域的核心任务之一,是计算机理解图像的基石。因此,增量目标检测问题具有重要的研究意义和实际意义。尽管增量学习在图像分类中取得了不错的成果,但基于目标检测的增量学习研究还处于初级阶段。这是因为目标检测相比图像分类更加复杂,它需要同时解决分类和边框回归的问题。不少研究者为解决此问题做了很多努力,但大多数工作都只关注如何保留已学习任务的性能,忽略了模型对新任务的快速适应能力,而这正是增量学习的关键要求。基于大脑的记忆机制,人类可以在学习中不断地提取知识以更好更快地学习新任务,不发生遗忘。受此启发,提出了一种融合编解码记忆重放机制的增量元学习方法。该方法对已学习样本的特征向量进行编码存储和解码重放,从而将不稳定数据流近似为动态稳定数据集,缓解了遗忘问题。同时,设计了一个双循环在线元学习策略,模型在内循环分别基于多批次新旧混合数据进行随机梯度更新,最后在外循环进行元学习,从而获得多任务间的共同结构,使模型具有良好的泛化性能,能够快速适应学习中遇到的新任务。在大型的公开数据集PASCAL VOC和MS COCO上设置了3种增量目标检测实验环境来评估所提算法。实验结果表明,所提算法与最先进的方法相比体现出了具有竞争力的性能,证明了其可以帮助模型更好地抵抗遗忘,具有更好的泛化性能。所提算法基于梯度更新,与模型无关,因此其可以与其他检测框架结合,具有强适应性。Incremental learning is key to bridging the enormous gap between artificial intelligence and human intelligence,mea-ning that agents can learn several tasks sequentially from a continuous stream of correlated data without forgetting,just as humans do.Object detection is one of the core tasks in the field of computer vision and the cornerstone of computer images understanding.Therefore,the incremental object detection has important research and practical significance.Although incremental learning has achieved good results in image classification,the research on incremental learning based on object detection is still in its infancy.This is because object detection is more complex than image classification,which needs to solve both classification and bounding box regression problems.Many researchers have made great efforts to solve this problem,but most of the work only focuses on how to retain previous learning,ignoring fast adaptability to new tasks,which is a critical requirement for incremental learning.Based on the memory mechanism of the brain,humans can constantly extract knowledge during learning,so as to learn new tasks better and faster without forgetting.Inspired by this,an incremental meta-learning method that integrates the codec memory replay mechanism is proposed.This method encodes,stores,decodes and replays the feature vectors of learned samples,so as to approximate the dynamic learning environment as a local stationary environment and avoid catastrophic forgetting.Besides,a double-loop online meta-learning strategy is designed,which can help model to extract common structures of tasks and improve generalization performance on new tasks encountered during learning.The model is respectively updated by SGD with multiple batches of old and new mixed data in the inner loop,and is meta-updated in the outer loop.We evaluate the proposed approach on three incremental object detection settings defined on PASCAL VOC and MS COCO datasets,where the proposed algorithm performs favorably well against state-of-th

关 键 词：增量学习 目标检测 受脑启发 元学习 抵抗遗忘 泛化性能 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]