Skip to content
All works
Sep 2026 – Present

FABRICPC

Research · SingularityNET open source

Building the benchmark suite for FabricPC, SingularityNET's open-source library for predictive coding, a brain-inspired alternative to backpropagation.

Architecture & approach

Predictive coding trains each layer on its own local error instead of one big backward pass. The open question is how it really compares to backprop, in accuracy and in compute cost. Our team is building fabricpc.bench so every result reproduces from one command, averaged over several seeds with proper statistics. Early runs: on MNIST, predictive coding matched backprop at about 98.2% accuracy across five seeds.