Hi, I'm the author (18yo high school dropout from Korea).
I built Versor because standard matrix multiplication (Wx+b) was warping the manifold of my physics data.
I needed a way to learn geometric transformations without fighting against the structure of the space.
The core idea: replace learned dense matrices with Clifford Algebra rotors (R = exp(-B/2)), which perform pure geometric rotations — isometries by construction. No stretching, no shearing, just alignment (I call this the "Unbending" paradigm).
In my Geometric Blade Network (GBN), I use what I call a Geometric FFT (GFFT) approach: approximating complex manifold transformations as a weighted superposition of simple rotors.
This reduces the parameter complexity from O(2^n) to O(n^2) while preserving physical laws.
Benchmarks:
QM9 (U0): 7.64 meV MAE in <1hr on a single 4090.
Efficiency: Real-time CPU inference at 5.8ms/molecule (leveraging high IPC/branch prediction of modern CPUs).
I originally built this for state estimation in my hardware project (Trajecto), where I needed a Tangent Space Observer to stable-track 3D trajectories.The code is Apache 2.0. I designed this more by geometric intuition than formal training. Please roast my code/math — I'm here to learn.
Hi, I'm the author (18yo high school dropout from Korea).
I built Versor because standard matrix multiplication (Wx+b) was warping the manifold of my physics data. I needed a way to learn geometric transformations without fighting against the structure of the space. The core idea: replace learned dense matrices with Clifford Algebra rotors (R = exp(-B/2)), which perform pure geometric rotations — isometries by construction. No stretching, no shearing, just alignment (I call this the "Unbending" paradigm). In my Geometric Blade Network (GBN), I use what I call a Geometric FFT (GFFT) approach: approximating complex manifold transformations as a weighted superposition of simple rotors. This reduces the parameter complexity from O(2^n) to O(n^2) while preserving physical laws. Benchmarks: QM9 (U0): 7.64 meV MAE in <1hr on a single 4090. Efficiency: Real-time CPU inference at 5.8ms/molecule (leveraging high IPC/branch prediction of modern CPUs).
I originally built this for state estimation in my hardware project (Trajecto), where I needed a Tangent Space Observer to stable-track 3D trajectories.The code is Apache 2.0. I designed this more by geometric intuition than formal training. Please roast my code/math — I'm here to learn.