Research papers and publication timeline.

NCG is a continual learning framework in which a network grows capacity dynamically—but only when it needs to. A novelty detector monitors incoming data and flags inputs that differ significantly from what the model has seen before; when novelty is detected, the network expands (new capacity) to accommodate new knowledge. Meta-parameters govern when and how much growth happens. The goal is to mitigate catastrophic forgetting—the loss of prior knowledge when training on new tasks.

In our paper we report 21% forgetting reduction on Split-MNIST (p = 0.012) and 64% on Split-CIFAR-10 (p < 0.0001). We also report an honest negative finding: meta-parameter recovery ratios sometimes fell below 0.5, meaning convergence is not always clean.

The NCG paper and release artifacts are now available with reproducible reporting and clear statements of both strengths and limitations.

ARIA is a more experimental continual learning direction, currently deprioritized relative to NCG. It targets a more ambitious architectural redesign than NCG’s capacity-growth framing.

Key components include sparse content routing (only relevant subnetworks activate for a given input, reducing interference between tasks), cortical columns (modular processing units inspired by neuroscience, for specialized knowledge domains), and Hebbian synaptic memory (connections strengthen from co-activation patterns). ARIA-Ω denotes a more advanced variant with further architectural changes.

The aria-torch package exists on PyPI, but the project is less mature than NCG and has not gone through the same rigorous experimental validation.

This track studies how well digital twins of real humans—trained on behavioral data—preserve fidelity in how each person thinks, reacts, and decides when composed into cohorts for simulation.

It supports Simfolk’s product direction (authentic twin populations versus generic personas) and is separate from DuckAI, which is an OCR-aware sidebar copilot—not a cohort of human-modeled twins.