A four-step loop that moves between neural and symbolic representations, designed to compress knowledge rather than scale parameters.
What it is
Artur d’Avila Garcez’s working definition of what makes a system neurosymbolic. The cycle has four steps:
- Translation — given a symbolic system, produce a corresponding neural network whose architecture or initial weights encode the symbolic knowledge
- Extraction — given a trained neural network, produce a symbolic description (rules, logic programs, decision trees, or graphs) that approximates the network’s behaviour
- Fidelity measurement — quantify how closely the extracted description matches the network using the fidelity metric
- Re-instillation — push the consolidated symbolic knowledge back into the network ahead of further training, closing the cycle
The slogan is learn a little, reason a little, repeat.
Why it matters
The cycle distinguishes neurosymbolic AI from systems that merely combine neural and symbolic components in an ad-hoc way. Each pass through the loop yields a smaller, more compressed network with externally-stated knowledge attached. This is the structural counter-thesis to the scale is all you need position: scaling NSAI means iterating the cycle, not adding parameters.
The bottleneck is consistently step 2 — extraction from large networks is hard, which is why the cycle is currently most viable on smaller or modular networks rather than on frontier-scale LLMs.