Sparse Hybrid Linear-Morphological Networks

Name: Sparse Hybrid Linear-Morphological Networks
Uploaded: 2026-10-15T20:05:59.341591+00:00
Duration: 197 s
Description: The method replaces standard network activation functions with explicitly sparse morphological layers to balance training and pruning. This hybrid structure uses max-plus operations, constrained by negative infinity initialization, resulting in better test performance on MTAT, faster convergence than dense models, and significantly enhanced pruning capabilities.

The method replaces standard network activation functions with explicitly sparse morphological layers to balance training and pruning. This hybrid structure uses max-plus operations, constrained by negative infinity initialization, resulting in better test performance on MTAT, faster convergence than dense models, and significantly enhanced pruning capabilities.

Open the full episode