Towards robust, interpretable neural networks via Hebbian/anti-Hebbian learning: A software framework for training with feature-based costs
Towards robust, interpretable neural networks via Hebbian/anti-Hebbian learning: A software framework for training with feature-based costs
复制标题
通过赫布/反赫布学习实现稳健、可解释的神经网络：基于特征成本的训练软件框架
DOI：
10.1016/j.simpa.2022.100347
复制
发表时间：
2022
期刊：
影响因子：
--
通讯作者：
Madhow, Upamanyu
中科院分区：
文献类型：
--
作者：
Cekic, Metehan;Bakiskan, Can;Madhow, Upamanyu
传统的基于端到端代价函数的深度神经网络（DNN）训练无法对深层神经网络各层提取的特征进行控制或提供保证。因此，尽管深度神经网络的影响普遍存在，但仍然存在对其（缺乏）可解释性和鲁棒性的重大担忧。在这项工作中，我们开发了一个软件框架，在这个框架中，端到端成本可以用依赖于分层激活的成本来补充，从而允许对功能进行更细粒度的控制。我们将该框架应用于判别设置中的Hebbian/anti-Hebbian （HaH）学习，展示了CIFAR10图像分类在鲁棒性方面的有希望的增益。
Conventional deep neural network (DNN) training with an end-to-end cost function is unable to exert control on, or to provide guarantees regarding the features extracted by the layers of a DNN. Thus, despite the pervasive impact of DNNs, there remain significant concerns regarding their (lack of) interpretability and robustness. In this work, we develop a software framework in which end-to-end costs can be supplemented with costs which depend on layer-wise activations, permitting more fine-grained control of features. We apply this framework to include Hebbian/anti-Hebbian (HaH) learning in a discriminative setting, demonstrating promising gains in robustness for CIFAR10 image classification.
影响因子：
56.9
作者：
Silver, David;Hubert, Thomas;Hassabis, Demis
通讯作者：
Hassabis, Demis
影响因子：
4.3
作者：
Burg MF;Cadena SA;Denfield GH;Walker EY;Tolias AS;Bethge M;Ecker AS
通讯作者：
Ecker AS