<ul><li>Deep State Space Models (SSMs) are bringing physics-grounded compute paradigms back into the spotlight.</li><li>Recurrent Hamiltonian Echo Learning (RHEL) is a new algorithm proposed to compute loss gradients efficiently for non-dissipative, Hamiltonian systems.</li><li>RHEL requires only three forward passes regardless of model size, without explicit Jacobian computation, ensuring consistent gradient estimation.</li><li>RHEL has been shown to match the performance of Backpropagation Through Time (BPTT) in training Hamiltonian SSMs on time-series tasks, demonstrating scalability and efficiency.</li></ul>

Learning long range dependencies through time reversal symmetry breaking

Discover more