Grokkit: A Unified Framework Grokked NN
https://doi.org/10.5281/zenodo.18072859
3 followers
https://doi.org/10.5281/zenodo.18072859
3 followers
Grokkit: Zero-shot resolution scaling for spectral operators I’ve released Grokkit, a framework for training neural networks as continuous operators rather than discrete functions. The core: Uses spectral consistency to enable zero-shot transfer across discretization scales. The result: Train on low-res, deploy on high-res with minimal MSE degradation (1.80 to 0.02). Efficiency: Deterministic, hallucination-free execution on consumer CPUs.
