This talk presents two wavelet-based innovations that significantly improve convolutional neural networks (CNNs) in terms of scalability, computational efficiency, and robustness. I’ll begin with WTConv, a novel convolutional layer that leverages multilevel Haar wavelet decomposition to expand receptive fields efficiently. WTConv scales logarithmically with kernel size, enabling nearly global receptive fields without the parameter bloat typical of large kernels. Beyond improving classification accuracy, WTConv boosts shape bias and robustness to corruptions, all while remaining a lightweight, drop-in replacement for depthwise convolutions.
Next, I’ll introduce Wavelet Compressed Convolution (WCC), a method for compressing high-resolution activation maps in image-to-image tasks. By applying joint wavelet-domain shrinkage across channels and executing 1×1 convolutions directly on the compressed representation, WCC substantially reduces memory bandwidth and compute cost. Unlike aggressive quantization—which often causes severe degradation—WCC maintains high accuracy across tasks like segmentation, super-resolution, and depth estimation, even under extreme compression. Together, these methods show how wavelet transforms can serve as a powerful, hardware-friendly toolset for designing scalable and efficient CNNs.
Bio:
Shahaf Finder is a PhD candidate in Computer Science, supervised by Prof. Oren Freifeld and Prof. Eran Treister, researching efficient neural networks with a focus on convolutional neural networks (CNNs) and graph neural networks (GNNs). He is also a Co-Founder of LimitlessCNC, a startup focused on integrating AI into CNC programming, where he leads the development of the core algorithmic technology.
