Official implementation of the paper: "Geometry-Aware Plasticity: Thermodynamic Weight Updates in Non-Euclidean Hardware" https://doi.org/10.5281/zenodo.18761137
GAP is a biologically plausible learning algorithm designed for Dynamically Gated Analog Crossbars (DGAC). It bridges the gap between the energy efficiency of local Hebbian learning and the global optimization power of backpropagation by utilizing dynamic Riemannian curvature.
-
The Hyperbolic Plunge: Exponentially scales learning rates along geodesic pathways during topological phase transitions.
-
The Hysteretic Shield: A Schmitt-trigger mechanism that immunizes learning against thermal noise.
-
Physical Governor: Enforces asymptotic memristive saturation bounds to prevent hardware burnout.
-
Asymmetric Gain: Protects baseline Euclidean memories during high-energy structural burns.
- Clone the repository:
git clone https://github.com/MPender08/Geometry-Aware-Plasticity.git
cd Geometry-Aware-Plasticity- Install dependencies:
pip install torch geoopt matplotlib numpy- Run the simulation:
python run_brain_sim.pyNGSpice is required to run the netlist (gap_learning.cir).
To generate the figure from the NGSpice data, make sure gap_learning_data.txt and plot_gap_data.py are in the same folder:
python plot_gap_data.pyFor the complete mathematical framework, please see the full manuscript included in this repository: geometry_aware_plasticity.pdf.
To see the Manifold Chip DGAC architecture, visit either the companion repository or the Zenodo page:
https://github.com/MPender08/manifold-chip-architecture
https://doi.org/10.5281/zenodo.18717807
To see the purely biological foundations of the Curvature Adaptation Hypothesis that inspired the DGAC hardware, visit my other companion repository: