在 VGG、U-Net、TCN 网络中... CNN 虽然功能强大,但必须针对特定问题、数据类型、长度和分辨率进行定制,才能发挥其作用。我们不禁会问,可以设计出一个在所有这些网络中都运行良好的单一 CNN 吗?
论文地址:https://arxiv.org/pdf/2206.03398.pdf 代码地址:https://github.com/david-knigge/ccnn
该研究提出 Continuous CNN(CCNN):一个简单、通用的 CNN,可以跨数据分辨率和维度使用,而不需要结构修改。CCNN 在序列 (1D)、视觉 (2D) 任务、以及不规则采样数据和测试时间分辨率变化的任务上超过 SOTA; 该研究对现有的 CCNN 方法提供了几种改进,使它们能够匹配当前 SOTA 方法,例如 S4。主要改进包括核生成器网络的初始化、卷积层修改以及 CNN 的整体结构。
![](https://image.jiqizhixin.com/uploads/editor/51dc9aef-db9f-4369-a96c-dda9c96466a5/640.png)
![](https://image.jiqizhixin.com/uploads/editor/110eee82-083d-4dc9-b599-c970974df9a3/640.png)
![](https://image.jiqizhixin.com/uploads/editor/0e04f894-6f9b-4046-8635-7caff6df3630/640.png)
![](https://image.jiqizhixin.com/uploads/editor/badc9cfd-a205-48b1-889b-9b16be6b99a7/640.png)
![](https://image.jiqizhixin.com/uploads/editor/82ac6932-2b5d-4189-a7e0-b5fe04e6b94c/640.png)
![](https://image.jiqizhixin.com/uploads/editor/80b914bf-bb68-450f-a7af-dc07dd982011/640.png)
![](https://image.jiqizhixin.com/uploads/editor/41daad72-5dfa-4905-9dc8-686189d87832/640.png)
![](https://image.jiqizhixin.com/uploads/editor/449cbd16-e0db-44f8-aae2-ac9845d7eb5f/640.png)
![](https://image.jiqizhixin.com/uploads/editor/a6e8932d-c98e-426c-bd6a-9c433ddb94df/640.png)
![](https://image.jiqizhixin.com/uploads/editor/de34df19-5edf-4b37-aaf3-1b5088222d11/640.png)
![](https://image.jiqizhixin.com/uploads/editor/8af23871-a2fc-47b6-b428-efc1411115d0/640.png)