🧠 Flexible Inference for Predictive Coding Networks in JAX ⚡️
📢 Updates
JPC is a JAX library for training neural
networks with Predictive Coding (PC).
JPC provides a simple, fast and flexible API for
training of a variety of PCNs including discriminative, generative and hybrid
models.
- Like JAX, JPC is completely functional in design, and the core library code is
<1000 lines of code.
- Unlike existing implementations, JPC provides a wide range of optimisers, both
discrete and continuous, to solve the inference dynamics of PC, including
ordinary differential equation (ODE) solvers.
- JPC also provides some analytical tools that can be used to study and
potentially diagnose issues with PCNs.
If you’re new to JPC, we recommend starting from the
example notebooks
and checking the documentation.
Overview
️💻 Installation
Clone the repo and in the project’s directory run
pip install .
Requires Python 3.10+ and JAX 0.4.38–0.5.2 (inclusive). For GPU usage, upgrade
jax to the appropriate cuda version (12 as an example here).
pip install --upgrade "jax[cuda12]"
Available at https://thebuckleylab.github.io/jpc/.
⚡️ Quick example
Use jpc.make_pc_step() to update the parameters of any neural network
compatible with PC updates (see the notebook examples
)
import jax.random as jr
import jax.numpy as jnp
import equinox as eqx
import optax
import jpc
# toy data
x = jnp.array([1., 1., 1.])
y = -x
# define model and optimiser
key = jr.PRNGKey(0)
model = jpc.make_mlp(
key,
input_dim=3,
width=50,
depth=5,
output_dim=3
act_fn="relu"
)
optim = optax.adam(1e-3)
opt_state = optim.init(
(eqx.filter(model, eqx.is_array), None)
)
# perform one training step with PC
result = jpc.make_pc_step(
model=model,
optim=optim,
opt_state=opt_state,
output=y,
input=x
)
# updated model and optimiser
model, opt_state = result["model"], result["opt_state"]
Under the hood, jpc.make_pc_step()
- integrates the inference (activity) dynamics using a diffrax ODE solver, and
- updates model parameters at the numerical solution of the activities with a given optax optimiser.
See the documentation for more details.
NOTE: All convenience training and test functions such as make_pc_step()
are already “jitted” (for optimised performance) for the user’s convenience.
🚀 Advanced usage
Advanced users can access all the underlying functions of jpc.make_pc_step()
as well as additional features. A custom PC training step looks like the
following:
import jpc
# 1. initialise activities with a feedforward pass
activities = jpc.init_activities_with_ffwd(model=model, input=x)
# 2. perform inference (state optimisation)
activity_opt_state = activity_optim.init(activities)
for _ in range(len(model)):
activity_update_result = jpc.update_pc_activities(
params=(model, None),
activities=activities,
optim=activity_optim,
opt_state=activity_opt_state,
output=y,
input=x
)
activities = activity_update_result["activities"]
activity_opt_state = activity_update_result["opt_state"]
# 3. update parameters at the activities' solution with PC
result = jpc.update_params(
params=(model, None),
activities=converged_activities,
optim=optim,
opt_state=opt_state,
output=y,
input=x
)
which can be embedded in a jitted function with any other additional
computations. Again, see the docs
for details.
Extensions
It would be good to add tooling for:
📄 Citation
If you found this library useful in your work, please cite (paper link):
@article{innocenti2024jpc,
title={JPC: Flexible Inference for Predictive Coding Networks in JAX},
author={Innocenti, Francesco and Kinghorn, Paul and Yun-Farmbrough, Will and Varona, Miguel De Llanza and Singh, Ryan and Buckley, Christopher L},
journal={arXiv preprint arXiv:2412.03676},
year={2024}
}
Also consider starring the repo! ⭐️
🙏 Acknowledgements
We are grateful to Patrick Kidger for early advice on how to use Diffrax.
See also: other PC libraries
评论