Scattering Networks for Hybrid Representation Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) are widely used for image modeling but often lack interpretability.
• Learned early layers in CNNs may not be essential for achieving high performance.

Purpose of the Study:

• To explore scattering networks as generic, interpretable representations for image modeling.
• To investigate the effectiveness of scattering networks in supervised and unsupervised learning tasks.
• To develop more interpretable CNN architectures.

Main Methods:

• Utilizing scattering networks with fixed weights as a replacement for early CNN layers.
• Developing hybrid architectures combining scattering networks with learned layers (e.g., 1x1 convolutions, residual networks).
• Applying scattering coefficients for image recovery in unsupervised learning and training Generative Adversarial Networks (GANs).

Main Results:

• Hybrid architectures achieved state-of-the-art results with predefined representations, competitive with end-to-end learned CNNs.
• A shallow scattering network matched AlexNet accuracy on the ILSVRC2012 classification task.
• Combining scattering networks with residual networks yielded a 11.4% top-5 error on ILSVRC2012.
• Scattering networks excelled in low-data regimes (CIFAR-10, STL-10) by incorporating geometric priors.
• Scattering coefficients proved effective for image recovery and training hybrid GANs.

Conclusions:

• Scattering networks provide a powerful and interpretable alternative to learned early layers in CNNs.
• Hybrid models integrating scattering networks offer significant performance gains and improved interpretability.
• Scattering networks demonstrate strong potential for both supervised and unsupervised image learning tasks, especially with limited data.