We just released a paper proposing a 'Topological Restructuring of AI'. Instead of training deep networks with Backpropagation (which is slow and energy-hungry), we project data onto a fixed Ramanujan Graph (optimal spectral expander) and solve the readout in closed-form.
Benchmarks:
Desktop (i5-4570, 4th gen): 287x speedup vs standard MLP training.
Mobile (Android ARM64): <0.6ms latency per inference (1600+ fps).
Accuracy: ~95.2% on MNIST (comparable to trained MLPs).
The code is C++ (for edge) and Python (for research). We believe this is a path towards deterministic, Green AI that runs on commodity hardware.Happy to answer questions about Spectral Graph Theory.