Computational Principles of Hierarchical Feature Emergence

Characterize the computational principles underlying the emergence of hierarchical features in biological neural systems, particularly in the cortex, to explain how such multi-level representations arise from sensory processing.

Background

The paper motivates biologically inspired architectures by noting that current AI models lack several key capabilities common in biological systems. A central challenge identified is understanding the mechanisms by which hierarchical features emerge in the brain, despite extensive empirical characterization of neural circuits.

This unresolved question frames the authors’ proposal of Rectified Spectral Units (ReSUs), which aim to learn predictive, hierarchical features via past–future canonical correlation analysis, offering a principled and biologically interpretable alternative to standard backpropagation-based deep networks.