Optimization

optimization is livn's differentiable ML layer containing metrics, losses, and fitting machinary used to estimate model parameters from data.

TIP

Naturally, this section assumes a differentiable environment requiring a JAX backend.

Installation

optimization ships with the repository as a workspace member:

uv sync --package optimization

It depends on livn[diffrax], so this also installs JAX, diffrax, equinox, and optax.

Hard and soft

Metrics come in two flavours:

• Hard metrics bin spike times onto a grid before comparing them. They are the numbers to report but the binning has no gradient, so they cannot be optimized against.
• Soft metrics are differentiable approximations over a continuous rate or raster. They are usable as fitting terms, and they are not expected to reproduce their hard counterpart's value.