Closed-form predictive coding via hierarchical Gaussian filters
A new paper presents a method for closed-form predictive coding using hierarchical Gaussian filters. This approach addresses the limitations of current predictive coding networks by incorporating precision-weighted message passing. The proposed method shows improved performance on various tasks compared to traditional backpropagation techniques.
- ▪Predictive coding offers a biologically grounded alternative to backpropagation in training neural networks.
- ▪The new method restores precision-weighted message passing, yielding dynamic uncertainty estimates.
- ▪The approach outperforms traditional methods on online, data efficiency, and concept-drift tasks.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.20293 (cs) [Submitted on 19 May 2026] Title:Closed-form predictive coding via hierarchical Gaussian filters Authors:Aleksandrs Baskakovs, Sylvain Estebe, Kenneth Enevoldsen, Kristoffer Nielbo, Chris Mathys, Nicolas Legrand View a PDF of the paper titled Closed-form predictive coding via hierarchical Gaussian filters, by Aleksandrs Baskakovs and Sylvain Estebe and Kenneth Enevoldsen and Kristoffer Nielbo and Chris Mathys and Nicolas Legrand View PDF Abstract:Predictive coding (PC) offers a local and biologically grounded alternative to backpropagation in the training of artificial neural networks, yet to date, it remains slower, and performance degrades sharply as network depth increases.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
