| Abstract: |
| The classical approximation theory developed 35 years ago is for fully-connected neural networks. This theory does not apply to neural networks with structures arising from applications of deep learning in speech recognition, computer vision, natural language processing, and many other domains. The structures and related network architectures raise some essential differences between the classical fully-connected neural networks and structured ones used in deep learning. This talk describes some approximation and generalization properties of structured neural networks such as deep convolutional neural networks. |
|