hyperbolix: Hyperbolic Deep Learning in JAX
The paper introduces an MIT-licensed JAX library meant to cover hyperbolic deep learning end to end.
`hyperbolix` is built on Flax NNX and presents a common interface across six manifolds, including Poincare ball, hyperboloid, mixed-curvature product spaces, and proper velocity space. The authors say it includes layers for linear models, convolutions, attention, normalization, positional encoding, regression, and vector quantization. It also ships Riemannian optimizers as optax transformations, wrapped distributions, and hyperbolic dimensionality-reduction tools. The paper highlights cancellation-free hyperboloid formulas that remain accurate in float32 at distances where prior implementations can return NaN. ArXiv · AI/CL/LG's note
`hyperbolix` is built on Flax NNX and presents a common interface across six manifolds, including Poincare ball, hyperboloid, mixed-curvature product spaces, and proper velocity space. The authors say it includes layers for linear models, convolutions, attention, normalization, positional encoding, regression, and vector quantization. It also ships Riemannian optimizers as optax transformations, wrapped distributions, and hyperbolic dimensionality-reduction tools. The paper highlights cancellation-free hyperboloid formulas that remain accurate in float32 at distances where prior implementations can return NaN. ArXiv · AI/CL/LG's note
score 5