# Composite-cycle interface search

- ID: 2875
- Canonical URL: https://synthcore.org/idea/2875/composite-cycle-interface-search
- API JSON: https://synthcore.org/api/idea/2875.json
- API Markdown: https://synthcore.org/api/idea/2875.md
- Verification status: unverified
- Source: [arXiv:2608.30231](https://arxiv.org/abs/2608.30231)
- Category: architecture
- Solves: accuracy, generalization, stability
- ML areas: rnn, ssm, architecture, curriculum
- Math tags: algebra, combinatorics, graph-theory, dynamical-systems
- Ratings: usefulness 6/10; difficulty 6/10; novelty 9/10

## Idea description

Use exhaustive finite-state analysis as an architecture-search objective for coupled recurrent motifs. Instead of independently tuning a normalization gate and a WTA gate, enumerate their possible interfaces and select couplings that create group action only on a small joint image set, yielding controlled reversible subdynamics embedded in an otherwise dissipative system.

## Mathematical statement

The paper distinguishes group structure inherited from an individual factor from group structure that exists only on the joint configuration space. In the coupled DN/WTA system, the pair C={(D:0,W:4),(D:1,W:5)} is an image set whose permutator group is Z_2, generated by exchanging the two composite states; the paper reports 35 composite group-carrying image sets out of 57, while the uncoupled product has no such composite components. Let S_D and S_W be finite state sets for two recurrent factors, let S=S_D times S_W, and let phi be an interface map controlling how the winner modifies normalization. For each candidate phi, construct generators f_u^phi on S, close them under composition, and score image subsets C on which some g in the monoid acts as a nontrivial permutation. A useful neural objective is to maximize composite-cycle score while penalizing cycles outside selected memory windows: L_interface = -N_comp(phi)+lambda N_unwanted(phi), where N_comp counts joint image sets with nontrivial permutation action not present in either factor alone.

## Key formulas

- $$C=\{(D{:}0,W{:}4),(D{:}1,W{:}5)\},\qquad \operatorname{Perm}(C)\cong\mathbb{Z}_2,$$
- $$S=S_D\times S_W,\qquad f_u^{\phi}(d,w)=\bigl(f_{D,u}(d,\phi(w)),\ f_{W,u}(w,d)\bigr),$$
- $$M_{\phi}=\langle f_u^{\phi}:u\in U\rangle,\qquad N_{\mathrm{comp}}(\phi)=\#\{C\subseteq S:\exists g\in M_{\phi},\ g(C)=C,\ g|_C\text{ is nonidentity and bijective}\},$$
- $$\mathcal{L}_{\mathrm{interface}}=-N_{\mathrm{comp}}(\phi)+\lambda\,\#\{C:\exists g,\ |C|>2,\ g|_C\text{ is a permutation}\}.$$

## Implementation notes

Integrate this as a discrete interface-search wrapper around a recurrent cell with two explicit finite factors. Use a small normalization state d in {0,...,D-1} and winner state w in {0,...,W-1}; the interface phi is a lookup table or a low-rank categorical map determining which normalization level is applied after each winner. For every candidate phi, enumerate the generator transition tables f_u^phi on S_D x S_W for a small input alphabet U, then perform monoid closure with a queue: start with the identity, compose each discovered transformation with every generator, hash transformation tables, and stop at closure or a fixed budget. For each subset encountered as an image of a monoid element, test whether any monoid element maps it bijectively to itself; classify the result as factor-local if the same permutation exists in a factor projection and composite otherwise. This is the paper-derived computation; gradients are not needed for the interface search. After selecting phi, replace lookup transitions with a differentiable implementation and train it on a temporal task. The cheap experiment is a 2-factor RNN on delayed XOR or parity with 4 normalization states, 4 winners, and 4 input symbols. Compare random coupling, uncoupled product, and algebraically selected coupling at equal parameter count. Measure task accuracy, robustness to state noise, the number and size of emergent cycles, and whether the selected interface improves long-horizon accuracy without causing global instability.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.