Cadence

Website · Cadence page · PyPI · Documentation

An experimental neural library for learning through local overlap repair and equilibrium detuning.

Cadence explores an animal-inspired hypothesis: the same network that interprets an observation should carry context, change its learned relationships through experience, and explore possible continuations. It does not claim to reproduce an animal or human brain.

Under active development. Pin a release or exact commit for reproducible work. Version 0.10.0 includes temporal learning, explicit response protection and a small equilibrium actor. Earlier applications use other library compositions; their results do not validate the revised architecture automatically.

Architecture and integration maps context, learning, explicit memory protection, self-readback, imagination and action planning to the implemented base-library APIs.

Design a task explains how observations, action ports, teaching and readback connect. Common missteps covers information loss, misleading proxy scores, memory interference and premature claims about coordination or scaling.

Temporal paths and persistent context

Version 0.10.0 adds experimental TemporalPatchNet. Each moment is an observer-like patch with input/output ports, bounded activity and a residual against its preceding state. During learning, adjacent patches repair a complete observed path; centered equilibrium detuning changes their shared relationships. Free inference carries hidden context, and private continuation uses the same learned relationships without changing live state. Targets never become the live hidden state.

The NumPy implementation exposes measured residuals, curvature checks, phase work and complete checkpoints. It extends the verified scalar-cue solver to time-varying observations. Sequence acquisition and long-term retention still require behavioral tests: a path can satisfy every model equation and predict poorly. This interface does not add a replay store or autonomous planning policy. The existing 0.9.0 graph API remains compatible.

TemporalMemory adds explicit response protection to temporal learning: caller-selected activity directions constrain later EP updates without replaying raw examples. Protected-path retention and remaining plasticity must be tested together; the available subspace is finite.

The base package also exposes EquilibriumActor: a minimal fixed-model observer plus joint future-state/action repair. It admits actual readings, retains a supplied goal, proposes an action privately and replans after real readback. Its linear Gaussian body assumptions and model-bound compressed history are explicit; it is not yet a general nonlinear composer.

Start with PatchNet

PatchNet is the existing general graph interface. It uses the existing nonlinear neural dynamics and local free/nudged learning rule, with a fully reciprocal graph by default. Its observer-like patches have bounded activity, declared ports, local readback and feedback/repair; experiments expose their observations, residuals and learned changes through reproducible evidence.

• Current context lives in neural activity. Activity continues between observations. Optional temporal overlap holds each solve against the previous free activity. Whether a particular graph retains a cue through a delay must be measured; a converged network can also forget its previous input.
• Acquired relationships live in continuous synapses and biases. Resetting activity leaves learned parameters intact. No external fact store or replay buffer is required by this interface. Interference during further learning remains a separate test.
• Real observations detune the network. Continuous targets nudge only declared observed ports. Each synapse changes from its endpoints' free/nudged activity contrast. The implementation does not construct a backward graph.
• Convergence is checked. Every required phase must satisfy the neural equations within the declared residual tolerance before learning commits. A capped solve is reported as unfinished. A small residual does not establish a unique, stable or correct answer.
• Imagined continuations are isolated. Branches use the same learned net without changing live activity, parameters or evidence bookkeeping. Branch isolation is implemented; useful planning and musical improvisation need empirical validation.
• A complete checkpoint resumes the learner. It includes parameters, optimizer state, current activity and the optional bounded source-ID window. Repeated IDs are suppressed within that window; IDs do not prove that two environmental reports are independent.

Follow the PatchNet guide for the running example, memory semantics, continuous targets and rehearsal. The earlier Brain, Learner, GenericBrain, Records and circuit APIs remain available for existing applications. Their separate associative memories are optional compositions, not required components of PatchNet.

The research target is fewer local mechanisms supporting acquisition, selective forgetting, retention and correction together. Learned importance, robust lifelong memory, autonomous specialization and animal-level capability remain open. In particular, the revised core does not freeze each weight into a binary state: that candidate prevented compatible learning through shared connections.

Install

Python 3.11+, with NumPy as the only required dependency:

python -m venv .venv
source .venv/bin/activate
python -m pip install cadence-net==0.10.0

For development from this checkout, use python -m pip install -e .. On Windows activate with .venv\Scripts\Activate.ps1. Optional backends support Numba, PyTorch and MLX.

Reference

Experience · Quickstart · Continuous interaction · Records · Write a cortex · Compose a brain · Evolve a brain · Local learning · Reward · API · All docs · Lean proofs

Check equation residuals before claiming equilibrium. Measure task quality and learning cost; local updates alone guarantee neither capability nor speed. Concepts and limits. MIT licensed.