Cadence

Website · Documentation · Examples · Paper · PyPI

An experimental brain that learns, remembers, imagines and acts.

Cadence aims to build a simulated human-like brain from simplified biological mechanisms. Its default System 1 is a continuing animal-like brain with perception, plastic connections, memory and action. Optional System 2 adds recursive feedback through observing cortical regions in the same neural graph. The base can already be deep and modular.

Bounded, observer-like regions carry local state, communicate through ports, read back activity and retain records. They repair disagreement to find a coherent state; actual observations and consequences guide learning. A settled answer can still be wrong about the world, so capability is measured through free behavior. Cadence is alpha research software, not a claim of human-level intelligence.

Animal and human brains learn from experience and not by backpropagation with gradient descent. Cadence is designed the same way. Each synapse changes from the activity of the two neurons it connects, compared between a free settled state and one nudged by the outcome. Reward scales that change, and memory writes are local too. The brain keeps no backward computation graph and has no separate training mode.

For embodied AI this design gives:

• Learning on the job. The same brain acts and learns from every measured outcome. There is no difference between training and inference.
• Adaptation to changed conditions. A changed body or world shows up in what the brain measures, and the live brain adjusts without being told what changed.
• Local learning. No backward pass through the network is needed, so learning can run where the brain runs.
• Memory. A working trace carries the recent past, and fast and persistent associative memory keep what mattered.
• Settled answers. Every action is a qualified settled state of the whole brain. A brain that does not settle refuses to act.
• Inspection. Region activity can be read while the brain runs, and private imagination tests a response before the body commits to it.

Under idealized conditions the local contrast follows the gradient that backpropagation would compute. The learning rule states those conditions, and the comparison is of update mechanisms. These properties are shown in simulation in the examples. An advantage on a physical robot is a separate test.

Python 3.11+ and NumPy are required.

python -m pip install cadence-net==0.70.0

import numpy as np
from cadence import Brain

brain = Brain.compose(inputs=4, actions=2, modules=(16, 8), seed=7)
observation = np.array([[1.0, 0.0, 0.0, 0.0]])
action = brain.step(observation)

# A tiny environment rewards action 0 and supplies the next observation.
reward = (action == 0).astype(float)
next_observation = np.array([[0.0, 1.0, 0.0, 0.0]])
action = brain.step(next_observation, reward=reward, done=np.array([False]))
assert action.shape == (1,)

step learns from the preceding action's measured reward, then chooses the next action. teacher= can label the current observation. Keep each batch row attached to the same life. There is no training/inference mode switch. Continuous interaction covers teaching, resets and saved continuation.

The constructor includes a working trace and fast/persistent associative memory. Earlier activity can affect later answers, and actual outcomes change associations. Capacity is finite; correlated memories can interfere.

phases = brain.imagine([observation, next_observation])
assert phases  # Inspect phase.converged before using an imagined response.

Imagination carries a private trace without changing live memory, random state or pending feedback. It evaluates responses to the observations you supply. For learned environmental consequences and action planning, use the separate temporal model.

recursive = Brain.compose(
    inputs=4, actions=2, modules=(16, 8), observers=(8,), seed=7,
)

Observer regions read and return influence to the base, motor regions and earlier observers. They join the same settlement and use the same interaction interface. This makes recursive feedback available; learning when it helps remains a task for experience and evaluation.

Actions and independent predictions require the full neural equation residual to meet the configured tolerance. Exhausting the budget refuses an action without changing its live state, memory or pending feedback. If step has learned a real outcome before the next action refuses, retry act without submitting that reward again. Numerical damping stays within the total budget and does not change the teaching rule. See contracts.

Build a brain for custom wiring, memory for traces and associations, and the memory/planning example for a bounded demonstration with actual toy-body outcomes. Record patches provide event records and consolidation. The advanced population solver provides exact state-and-error readback under its own numerical contract.

Documentation · API · Contributing · Changelog · Research tasks

Licensed under GPL-3.0.