Code for Shared Phase and Retention Control for Efficient Adaptive Spectral Recurrence.
- JAX reference implementation: Flax layers, actor-critic and sequence-classification networks, structured RTRL and BPTT training, and usage examples.
- GPU-efficient implementation: a differentiable PyTorch module with Triton forward and backward operators for sequence training and single-step decoding.
See each implementation's README for dependencies, installation, and usage.