FabricPC Roadmap
Contents
Making Predictive Coding Accessible, Scalable and Reproducible
Since FabricPC was officially introduced in June, several new capabilities have been added to the framework. These developments and planned improvements fall into three main areas:
- Accessibility: Lowering the barriers to using FabricPC by simplifying training, installation, documentation, and other parts of the research workflow.
- Scalability: Supporting deeper predictive coding networks while reducing the computational cost required to train them.
- Reproducibility: Building tools that compare methods fairly and let others reproduce the results.
This article reviews those developments and outlines the next steps on the roadmap. For the sequence, current status, and technical details of individual components, see the FabricPC roadmap on GitHub.
What FabricPC Is
FabricPC is an open-source, high-performance Python library for building and training neural networks with predictive coding. Predictive coding allows different parts of a network to learn and adjust based on locally available information. This makes it an interesting research direction for systems that keep learning over time and for building more flexible neural architectures. One potential application is enabling models such as LLMs to incorporate new information without retraining from scratch.
A deeper introduction and overview of its existing capabilities is available in the introductory article, while the open-source library, documentation, and demos are available through the GitHub repository.
Accessibility
Making predictive coding easier to work with has been a guiding principle of FabricPC from the beginning. Recent development added a unified training interface that brings different training configurations into one consistent workflow, reducing duplicated setup when researchers compare different approaches.
The roadmap extends that work across more of the research process. Model checkpointing will allow experiments to be saved and resumed from the same state while supporting pretrained models and reproducible starting points for further work. Shorter example configurations, a Colab quickstart that removes the need for a local installation, and continued improvements to documentation and demos will further reduce the effort required to get started and run experiments.
Scalability
Accessibility also depends on researchers being able to work with advances from across the predictive coding field without moving between paper- or lab-specific implementations. FabricPC brings new advances into a shared environment that can be used across different experiments and architectures.
One example is muPC, an approach for scaling predictive coding across greater network depth and width. FabricPC incorporated muPC and extended its original formulation beyond sequential chains to more flexible graph structures.
This week’s v0.6.0 release added error-based predictive coding, or ePC, as another method for scaling predictive coding to deeper networks. FabricPC already supported state-based predictive coding (sPC). Both are inference solvers that determine how the network adjusts its internal state before learning takes place.
With sPC, the network repeatedly adjusts the internal states of its layers, which in turn changes the prediction errors. ePC reverses that relationship by working directly with the prediction errors, which then determine how the internal states should change. This allows useful learning signals to propagate through deeper networks with fewer repeated inference steps, making deeper predictive coding networks more practical to train while reducing computational cost.
The same release introduced composable inference solvers, allowing ePC and sPC to be chained within the same inference process, and extended ePC to cyclic graphs through cycle unrolling.
Further research will focus on improving ePC stability during predictive coding training. Work will also continue to reduce compilation overhead in larger graphs and make iterative refinement more selective. But determining how best to use these new capabilities across different architectures will require systematic empirical evaluation.
Reproducibility
FabricPC already provides an environment for comparing predictive coding with backpropagation. The same graph architecture can be trained with either approach, while its experiment framework supports repeated trials and statistical comparison of results.
The roadmap extends this foundation with a broader benchmark suite for systematically evaluating predictive coding methods across deeper networks, different inference approaches, and cyclic architectures. Documented configurations and versioned results will provide a clear path for reproducing results, while additional validation tooling and pretrained reference models will make established results easier to verify, inspect, and build on.
Where This Leads
Predictive coding addresses questions that are directly relevant to the broader AGI research pursued across the ASI Alliance and SingularityNET. How can neural networks continue learning from new information without repeatedly retraining the entire system? How can they adapt over long periods as they encounter new information and changing environments? These are among the fundamental questions on the path toward general intelligence.
FabricPC contributes to this effort by making predictive coding easier to experiment with in practice. Lowering the barriers to testing, comparing, and extending these approaches can bring more researchers into the field, enable more ideas to be explored, and accelerate predictive coding research toward more capable and adaptive AI systems.
FabricPC is fully open-source. Researchers and developers can use it for their own experiments, explore the documentation and demos, or inspect the development backlog and contribute through issues and pull requests.
