Reproduce “Neural Field Turing Machine: A Differentiable Spatial Computer” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Akash Malhotra, Nacéra Seghouani
The Neural Field Turing Machine (NFTM) repeatedly updates a spatial field using local information. Its controller reads local regions and proposes updates; the architecture can also move read/write heads. In CIFAR-10 inpainting, a model trained for 20 rollout steps improved PSNR from 15.2 dB at step 1 to 24.5 dB at step 30, showing refinement past training, not beating baselines.
Key insights
- NFTM is formally proven to be Turing complete under bounded error, establishing its fundamental expressive power as a general-purpose computational model for continuous spatial domains.
- The architecture's local update rules, operating on continuous fields with fixed-radius neighborhoods, result in a linear computational cost (O(N)) per timestep, ensuring efficiency for large spatial problems.
- NFTM demonstrates adaptive computation, where accuracy improves (e.g., PSNR increase in inpainting) when rolled out for more steps than during training, indicating it learns a stable iterative refinement process.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.