MSc Thesis: Dynamic 3D Networks

The attached file exhibits an amended version of the author's MSc thesis (originally submitted - September 2024). A 3D spatially-embedded neural network has parameters conditioned on node coordinates. The network's structure entirely determines its function. The architecture trains via a reinforcement update that transforms a scalar field on R^3. The model was trained on the MNIST image classification dataset, but produced uncompetitive results. Embedding in higher and lower-dimensional spaces is discussed. The author also submitted the project to the ICL Computing department's admissions panel (November 2024).