1: \begin{abstract}

2: Convolutional neural networks (CNNs) have enabled the state-of-the-art performance in many computer vision tasks.

3: However, little effort has been devoted to establishing convolution in non-linear space.

4: Existing works mainly leverage on the activation layers, which can only provide point-wise non-linearity.

5: To solve this problem, a new operation, kervolution (\textbf{ker}nel con\textbf{volution}), is introduced to approximate complex behaviors of human perception systems leveraging on the kernel trick.

6: It generalizes convolution, enhances the model capacity, and captures higher order interactions of features, via patch-wise kernel functions, but without introducing additional parameters.

7: Extensive experiments show that kervolutional neural networks (KNN) achieve higher accuracy and faster convergence than baseline CNN.

8: \end{abstract}

9: