Abstract:Continuous convolution has recently gained prominence due to its ability to handle irregularly sampled data and model long-term dependency. Also, the promising experimental results of using large convolutional kernels have catalyzed the development of continuous convolution since they can construct large kernels very efficiently. Leveraging neural networks, more specifically multilayer perceptrons (MLPs), is by far the most prevalent approach to implementing continuous convolution. However, there are a few drawbacks, such as high computational costs, complex hyperparameter tuning, and limited descriptive power of filters. This paper suggests an alternative approach to building a continuous convolution without neural networks, resulting in more computationally efficient and improved performance. We present self-moving point representations where weight parameters freely move, and interpolation schemes are used to implement continuous functions. When applied to construct convolutional kernels, the experimental results have shown improved performance with drop-in replacement in the existing frameworks. Due to its lightweight structure, we are first to demonstrate the effectiveness of continuous convolution in a large-scale setting, e.g., ImageNet, presenting the improvements over the prior arts. Our code is available on <a class="link-external link-https" href="https://github.com/sangnekim/SMPConv" rel="external noopener nofollow">this https URL</a>
What problem does this paper attempt to address?
The problems that this paper attempts to solve mainly focus on the efficient implementation of continuous convolution. Specifically, the paper aims to propose a method that does not rely on neural networks to construct continuous convolution kernels, in order to reduce computational costs and improve performance. Traditional methods usually use multi - layer perceptrons (MLP) to implement continuous convolution, but this method has disadvantages such as high computational costs, complex hyper - parameter tuning, and limited filter description capabilities. Therefore, this paper proposes a new method - Self - moving Point Representations (SMP), which allows weight parameters to move freely and uses an interpolation scheme to implement continuous functions, thereby improving the performance of continuous convolution while maintaining a lightweight structure.
### Main contributions
1. **Propose SMPConv**: A continuous convolution method based on self - moving point representation, which does not require the use of neural networks, thereby reducing computational costs and improving performance.
2. **Efficient construction of large - scale convolution kernels**: SMPConv can effectively construct large - scale convolution kernels on large - scale datasets (such as ImageNet), which was difficult to achieve with previous methods.
3. **Extensive experimental verification**: Through experiments on multiple tasks, including continuous function approximation, sequence data classification, image classification, etc., the effectiveness and superiority of SMPConv have been verified.
### Specific problems solved
- **High computational cost**: Traditional methods use MLP to implement continuous convolution, resulting in high computational costs, especially when dealing with large - scale datasets.
- **Complex hyper - parameter tuning**: The MLP method requires complex hyper - parameter tuning, increasing the difficulty of model training.
- **Limited filter description capabilities**: The convolution kernels generated by traditional methods have limited description capabilities for high - frequency components, affecting the performance of the model.
### Experimental results
- **Continuous function approximation**: Experiments show that SMPConv can accurately approximate continuous functions with a smaller number of points.
- **Sequence data classification**: In sequence image and time - series classification tasks, SMPConv achieves better performance than existing methods.
- **Image classification**: On the CIFAR10 dataset, SMPConv has better performance than traditional methods such as ResNet with a smaller number of parameters.
- **Large - scale image classification**: On the ImageNet dataset, SMPConv achieves performance comparable to or even better than existing methods with a smaller number of parameters.
In conclusion, this paper solves the key problems of continuous convolution in computational efficiency and performance by proposing SMPConv, providing a new solution for continuous convolution applications on large - scale datasets.