The convolutional neural networks (CNNs) are a powerful tool of image classification that has been widely adopted in applications of automated scene segmentation and identification. However, the mechanisms underlying CNN image classification remain to be elucidated. In this study, we developed a new approach to address this issue by investigating transfer of learning in representative CNNs (AlexNet, VGG, ResNet-101, and Inception-ResNet-v2) on classifying geometric shapes based on local/global features or invariants. While the local features are based on simple components, such as orientation of line segment or whether two lines are parallel, the global features are based on the whole object such as whether an object has a hole or whether an object is inside of another object. Six experiments were conducted to test two hypotheses on CNN shape classification. The first hypothesis is that transfer of learning based on local features is higher than transfer of learning based on global features. The second hypothesis is that the CNNs with more layers and advanced architectures have higher transfer of learning based global features. The first two experiments examined how the CNNs transferred learning of discriminating local features (square, rectangle, trapezoid, and parallelogram). The other four experiments examined how the CNNs transferred learning of discriminating global features (presence of a hole, connectivity, and inside/outside relationship). While the CNNs exhibited robust learning on classifying shapes, transfer of learning varied from task to task, and model to model. The results rejected both hypotheses. First, some CNNs exhibited lower transfer of learning based on local features than that based on global features. Second the advanced CNNs exhibited lower transfer of learning on global features than that of the earlier models. Among the tested geometric features, we found that learning of discriminating inside/outside relationship was the most difficult to be transferred, indicating an effective benchmark to develop future CNNs. In contrast to the “ImageNet” approach that employs natural images to train and analyze the CNNs, the results show proof of concept for the “ShapeNet” approach that employs well-defined geometric shapes to elucidate the strengths and limitations of the computation in CNN image classification. This “ShapeNet” approach will also provide insights into understanding visual information processing the primate visual systems.

卷积神经网络（CNN）是一种强大的图像分类工具，已在自动场景分割和识别应用中广泛采用。但是，CNN图像分类的基础机制仍有待阐明。在这项研究中，我们研究了代表性CNN（AlexNet，VGG，ResNet-101和Inception-ResNet-v2）中基于局部/全局特征或不变量对几何形状进行分类的学习方法，从而开发了一种解决此问题的新方法。局部特征基于简单的成分（例如线段的方向或两条线是否平行）而全局特征则基于整个对象，例如对象是否有孔或对象是否在另一个对象内部。进行了六个实验，以检验关于CNN形状分类的两个假设。第一个假设是，基于局部特征的学习转移高于基于全局特征的学习转移。第二个假设是，具有更多层和高级架构的CNN具有更高的基于学习的全局特征转移。前两个实验检查了CNN如何转移识别局部特征（正方形，矩形，梯形和平行四边形）的学习。其他四个实验检查了CNN如何转移识别全局特征（孔的存在，连通性以及内部/外部关系）的学习。尽管CNN在分类形状方面表现出强大的学习能力，但学习的转移因任务而异，并且因模型而异。结果拒绝了两个假设。第一的，一些CNN展示的基于本地特征的学习转移低于基于全局特征的学习转移。其次，先进的CNN与早期模型相比，在全球功能方面的学习转移较少。在经过测试的几何特征中，我们发现区分内部/外部关系的学习最为困难，这表明了开发未来CNN的有效基准。与采用自然图像训练和分析CNN的“ ImageNet”方法相比，结果显示了“ ShapeNet”方法的概念证明，该方法使用定义明确的几何形状来阐明CNN图像中计算的优势和局限性分类。这种“ ShapeNet”方法还将提供洞察力，帮助您了解处理灵长类动物视觉系统的视觉信息。