Training sparse convolutional deep predictive coding networks with attention

doi:10.1016/j.neunet.2026.108836

Training sparse convolutional deep predictive coding networks with attention

2026-03-12

0

PRE

AI

DOI：10.1016/j.neunet.2026.108836

Original

Original request for help

Save

Abstract

En 中文

• The paper proposes a novel deep predictive coding network (DPCN) learning algorithm that improves upon previous DPCNs by incorporating during the training phase a bidirectional structure with top-down attention to improve the quality of the internal representations, which has been an unresolved issue. • Consisting of six convolutional sparse coding layers, the proposed model efficiently preserves object-level features in the deepest layer, which remain visually interpretable even with high sparsity (less than 1% to 5% non-zero values). • We also propose a new visualization method applicable to the deep layers of DPCN as well as other sparse coding approaches. This method projects activations from deep receptive fields back onto the input space, enabling a clear observation of the attention patterns associated with individual learned convolution filters.

Keywords:

Deep Predictive Coding Network

Sparse Coding

Attention Mechanism

Convolutional Neural Networks

Visual Interpretability

Training sparse convolutional deep predictive coding networks with attention

Abstract

Journal

Organization