Efficient Hybrid Network: Inducting Scattering Features
This work addresses a core limitation in hybrid networks for machine learning applications, offering improved performance in data-rich scenarios while maintaining generalization in data-limited settings.
The paper tackles the performance gap of hybrid networks compared to conventional ones when ample training data is available, introducing Efficient Hybrid Network (E-HybridNet) which consistently outperforms conventional counterparts on diverse datasets.
Recent work showed that hybrid networks, which combine predefined and learnt filters within a single architecture, are more amenable to theoretical analysis and less prone to overfitting in data-limited scenarios. However, their performance has yet to prove competitive against the conventional counterparts when sufficient amounts of training data are available. In an attempt to address this core limitation of current hybrid networks, we introduce an Efficient Hybrid Network (E-HybridNet). We show that it is the first scattering based approach that consistently outperforms its conventional counterparts on a diverse range of datasets. It is achieved with a novel inductive architecture that embeds scattering features into the network flow using Hybrid Fusion Blocks. We also demonstrate that the proposed design inherits the key property of prior hybrid networks -- an effective generalisation in data-limited scenarios. Our approach successfully combines the best of the two worlds: flexibility and power of learnt features and stability and predictability of scattering representations.