Cadence

Website · Demo · Paper · PyPI · Documentation

Build Deep Recursive Settlement Networks: processing populations and their observers, learning and settling together.

Cadence explores a simple idea: a brain can be a persistent system of local relations that repairs its state as experience arrives. An observer reads what other patches are doing and influences them while remaining part of the same equilibrium. Another observer can read that enlarged system in turn.

The aim is a straightforward replacement for legacy deep neural networks, including transformers, built around persistent recursive settlement. Performance comparisons are ongoing.

Cadence implements this population architecture with a standard-library Python engine and optional tensor execution on CPU, Apple Silicon GPU or NVIDIA GPU. Every processing and observer patch uses one local prediction relation. The compiled brain supports joint settlement, supervised witness learning, diagnostics and continuation.

In the bootstrapping phase, guided experience teaches the brain its basic abilities. In the live phase, that same brain acts in its environment and can continue learning from actual new witnesses. Both phases use the same patch rule and retained world model.

Three pillars

• Minimalism: one patch rule, a small public API, no required runtime dependencies. Focused modules keep each responsibility easy to find.
• User-friendliness: explicit layouts, sensible defaults and runnable examples.
• Agent-friendliness: exact contracts, inspectable state and explicit refusal.

Every core change is reviewed against all three. The contributor instructions define the checks.

The demo: a brain plays Doom

The main Cadence demo is Doom Lab: a settling brain plays shareware Doom in your browser from pixels alone — two grayscale retinas in, eight settled motor states out, no policy network and no backpropagation. Take the controls yourself and your frame-and-button pairs are admitted as witnesses while you play; a live, rotatable 3D view shows the populations settling, the wiring carrying signal, and each committed batch as it changes the brain. Brains hot-swap mid-game and travel as single export files.

git clone https://github.com/muellerberndt/cadence-demos
cd cadence-demos/doom-lab && ./get_wad.sh && python server.py

The complete training pipeline ships with it — scripted privileged teacher, witness corpora, conditioned batched bootstrap with matched controls, DAgger correction rounds, and a foresight variant trained on measured outcomes: demo page · training tutorial.

Install

Python 3.11 or later; the default engine has no runtime dependencies:

python -m pip install --upgrade "cadence-net>=0.49.0"

If your package index has not listed a new release yet, its wheel is also available from GitHub Releases.

For GPU execution, install the optional tensor extra:

python -m pip install --upgrade "cadence-net[gpu]"

Choose Cortex(device="mps") for an Apple Silicon GPU or Cortex(device="cuda") for a compatible NVIDIA GPU. The same brain methods and learning rule apply. Final admission is checked against the original float64 objective. Small brains can be faster with the default Python engine; see GPU execution and parallel experience for precision, measurement and independent simulated lives.

Declare the brain

from cadence import Cortex

cortex = Cortex(seed=7)
eyes = cortex.input("eyes", shape=(8, 8))
ears = cortex.input("ears", shape=(2, 16))
body = cortex.input("sensory_nerves", shape=(8,))
senses = (eyes, ears, body)

c1 = cortex.column("perception", patches=16, inputs=senses)
c2 = cortex.observer(
    "integration", patches=8, inputs=senses, observes=(c1,),
)
c3 = cortex.observer(
    "reflection", patches=8, inputs=senses, observes=(c1, c2),
)
cortex.output("motor_nerves", shape=(8,), reads=c3)
brain = cortex.build()

assert brain.inspect()["patches"] == 32
assert brain.inspect()["sensor_coverage"] == 104

inputs supplies sensory or represented data. observes adds readback of live patch states and exact current prediction errors. Observer constraints feed back into the observed states through the same energy and repair process. They are not a separate decision made after the lower system finishes.

Population size counts processing patches. Sensor shape counts input samples; a larger camera shape does not supply learned visual understanding. The eight outputs expose eight settled patch values, without a separate policy network. The guide also shows parallel sensory branches and inspection of actual wiring. Every patch reads all coordinates of its declared sources by default; explicit fan_in opts into sparse sampling. This is a layout example, not a pretrained vision/audio model. For a first learning task, start smaller using the bootstrapping guide. Inspection also reports sensor coverage for each output coordinate, so you can check the decision's connections before bootstrapping.

Bootstrapping phase

This inexpensive example bootstraps a small brain's signed input/output relation. It uses the same population primitive, with fewer patches for a quick run.

from cadence import bootstrap

bootstrap_layout = Cortex(seed=2)
signal = bootstrap_layout.input("signal", shape=(1,))
base = bootstrap_layout.column("base", patches=4, inputs=signal)
reflection = bootstrap_layout.observer(
    "reflection", patches=2, inputs=signal, observes=base,
)
bootstrap_layout.output("answer", shape=(1,), reads=reflection)
learner = bootstrap_layout.build()

examples = [({"signal": [x]}, {"answer": [x]}) for x in (-0.8, 0.8)]
checks = [({"signal": [x]}, {"answer": [x]}) for x in (-0.4, 0.4)]
report = bootstrap(learner, examples, checks=checks, max_error=0.2)
assert report["passed"], report

# Fresh amplitudes, with no output targets supplied to the brain.
assert learner.predict({"signal": [-0.4]})["answer"][0] < -0.2
assert learner.predict({"signal": [0.4]})["answer"][0] > 0.2

Actual witnesses clamp the supplied outputs and allow joint repair of live state and retained relation parameters. Only qualified proposals are committed. Queries hold learned parameters fixed; hypothetical clamps never become experience. Learning is evaluated on later unclamped predictions, as above. This example establishes a small acquired relation, not a benefit from depth.

Use batch bootstrapping to jointly learn from groups of examples:

batch_report = bootstrap(
    learner, examples, checks=checks, max_error=0.2, batch_size=32,
)
assert batch_report["passed"], batch_report

Each batch jointly settles private example states and shared parameters; CPU/GPU tensors vectorize the examples. brain.observe_batch(examples) admits one batch directly. Batch commits preserve live activity and update parameters atomically. The default batch_size=1 keeps ordered single-experience learning. See batch bootstrapping for the objective and resource tradeoffs.

Live phase

activity = learner.step({"signal": [0.3]})
assert activity["accepted"]
print(activity["outputs"]["answer"])

# When an actual teaching signal arrives, the same brain can keep learning.
assert learner.observe({"signal": [0.3]}, {"answer": [0.3]})["accepted"]

bootstrap supplies reproducible replay and unclamped readiness checks. Applications supply sensory acquisition, witnesses, actuator interpretation and environment-specific behavior checks. The guide explains calibration and staged bootstrapping for perception and body control.

Development support for a continuing life adds explicit sensory History, discrete reward learning with Reinforcement, an optional LearningProgress curiosity signal, and a LiveController that keeps rendering independent of brain work. Reward-derived action values are labeled estimates and learned by the same patch repair. Replay and checkpointing preserve experience for later use. These additions follow 0.49.0; see memory, rewards and live control for setup, examples and the boundary between learned decisions and body support.

Scope

The reference engine repairs a nonlinear residual energy using analytic derivatives and a bounded descent procedure. Every participating population is included in the final constrained-stationarity check. Prediction errors can remain nonzero at a qualified compromise; different starting states can reach different stationary points. A solver can refuse when its budget is exhausted.

Performance evaluations and comparisons remain ongoing. Tests cover layouts, derivatives through recursive error readback, reciprocal interventions, actual acquisition, refusal and checkpoint custody. The goal is a reusable learner across perception, memory, reasoning and embodied action. Discrete Q-learning and bounded temporal history have small behavioral checks; long-lived learned memory, broader capability and advantages from recursive depth require further controlled experiments. Follow the DRSN completion epic.

GPL-3.0-or-later license.