摘要
人工智能使人类改造自然、适应自然的各类技术发展到更高阶段,是人类社会发展过程中的一次重要革命.深度学习不可解释性难题是当前人工智能领域的瓶颈问题.基于深度模型的推理过程是一个黑箱,现有的理论还不能完全解释模型输出结果的原因,对它的研究还处于比较初级的阶段.针对双线性模型的不可解释问题,我们提出了基于协同梯度下降算法的可信学习方法,在模型优化过程中引入反馈机制实现模型解耦,构建了可信的因果模型,提升了模型的性能和可解释性.在卷积神经网络训练以及模型压缩的任务中,实验结果表明该方法的有效性和适用性.
Artificial intelligence enables human beings to transform and adapt to nature to a higher stage, which is a revolution in thedevelopment of human society. However, the poor interpretability of deep learning is currently the bottleneck for artificialintelligence. The reasoning process based on the deep model is a black box, and the existing theories cannot fully explain the reasonsfor the output results of the model, and the research on it is still at a relatively early stage. To improve the interpretability of bilinearmodels, we put forward a dependable learning approach based on the cogradient descent algorithm. A feedback mechanism isintroduced in the training process to realize decoupling and causality consolidation, which improves the interpretability andperformance of the model. In the tasks of convolutional neural network training and model compression, the results prove theeffectiveness and applicability of the method.
作者
张宝昌
鲍宇翔
王润琪
吕金虎
刘克新
ZHANG BaoChang;BAO YuXiang;WANG RunQi;LÜJinHu;LIU KeXin(Institute of Artificial Intelligence,Beihang University,Beijing 100191,China;School of Automation Science and Electrical Engineering,Beihang University,Beijing 100191,China)
出处
《中国科学：技术科学》
EI
CSCD
北大核心
2024年第2期257-264,共8页
Scientia Sinica(Technologica)
基金
国家自然科学基金(批准号:62076016)。
关键词
可信学习
双线性优化
因果关系
协同梯度下降
解耦
dependable learning
bilinear optimization
causality
cogradient descent
decoupling
作者简介
张宝昌,E-mail:bczhang@buaa.edu.cn。
