Search papers, labs, and topics across Lattice.
The paper introduces the Causal Hamiltonian Learning Unit (CHLU), a novel deep learning primitive grounded in physics, designed to address the stability-memory trade-off in temporal dynamics modeling. CHLUs enforce a Relativistic Hamiltonian structure and use symplectic integration to strictly conserve phase-space volume, aiming for infinite-horizon stability and controllable noise filtering. The authors demonstrate the generative capabilities of the CHLU on the MNIST dataset, providing a proof-of-principle for its effectiveness.
Escape the memory-stability trade-off in temporal models with CHLUs, a new primitive that uses Hamiltonian structure and symplectic integration to conserve information over time.
Current deep learning primitives dealing with temporal dynamics suffer from a fundamental dichotomy: they are either discrete and unstable (LSTMs) \citep{pascanu_difficulty_2013}, leading to exploding or vanishing gradients; or they are continuous and dissipative (Neural ODEs) \citep{dupont_augmented_2019}, which destroy information over time to ensure stability. We propose the \textbf{Causal Hamiltonian Learning Unit} (pronounced: \textit{clue}), a novel Physics-grounded computational learning primitive. By enforcing a Relativistic Hamiltonian structure and utilizing symplectic integration, a CHLU strictly conserves phase-space volume, as an attempt to solve the memory-stability trade-off. We show that the CHLU is designed for infinite-horizon stability, as well as controllable noise filtering. We then demonstrate a CHLU's generative ability using the MNIST dataset as a proof-of-principle.
