# Criticality-Gated Resolution Switching

- ID: 761
- Canonical URL: https://synthcore.org/idea/761/criticality-gated-resolution-switching
- API JSON: https://synthcore.org/api/idea/761.json
- API Markdown: https://synthcore.org/api/idea/761.md
- Verification status: unverified
- Source: [arXiv:2607.11661](https://arxiv.org/abs/2607.11661)
- Category: dynamics
- Solves: stability, accuracy, speedup
- ML areas: graph-nn, cnn, transformer, inference-speedup
- Math tags: statistical-mechanics, renormalization, bifurcations, dynamical-systems
- Ratings: usefulness 5/10; difficulty 6/10; novelty 8/10

## Idea description

Use effective coupling and field values from a local coarse-grained motif to decide whether a neural network should operate at fine or coarse resolution. Near the continuous critical boundary, retain fine-scale features because correlations become long-ranged; away from criticality, aggregate aggressively. Near discontinuous or reentrant boundaries, hysteresis prevents rapid switching between resolutions.

## Mathematical statement

For the effective zero-field square-lattice Ising model, the continuous transition occurs at the dimensionless coupling \(K_c=\frac12\log(1+\sqrt2)\), where \(K_{\mathrm{eff}}=\beta J_{\mathrm{eff}}\). Define the criticality score \(r=|K_{\mathrm{eff}}-K_c|\) and symmetry score \(q=|h_{\mathrm{eff}}|\). The two-dimensional Ising correlation length scales as \(\xi\sim |K_{\mathrm{eff}}-K_c|^{-\nu}\), with critical exponent \(\nu=1\). A resolution controller selects fine processing when \(r\) is below an activation threshold or when the effective field is sufficiently small. Hysteresis uses an activation threshold \(\delta_{\mathrm{on}}\) smaller than a deactivation threshold \(\delta_{\mathrm{off}}\), preventing chattering.

## Key formulas

- $$K_c=\frac{1}{2}\log(1+\sqrt{2}),\qquad r=|K_{\mathrm{eff}}-K_c|,\qquad q=|h_{\mathrm{eff}}|$$
- $$\xi\propto |K_{\mathrm{eff}}-K_c|^{-\nu},\qquad \nu=1$$
- $$g_{t+1}=\begin{cases}1,&r_t<\delta_{\mathrm{on}}\ \text{or}\ q_t<\delta_h\\0,&r_t>\delta_{\mathrm{off}}\ \text{and}\ q_t\geq\delta_h\end{cases}$$

## Implementation notes

1. Integration point: add a controller to a hierarchical GNN, multiscale CNN, or token-merging transformer. Each local patch or graph motif produces \(K_{\mathrm{eff}}\) and \(h_{\mathrm{eff}}\); these values control whether its fine tokens or nodes are retained or merged into one coarse token. 2. Pseudocode: for each patch compute the three local weights by summing over the auxiliary state; calculate `Keff` and `heff`; set `fine=True` if `abs(Keff-Kc)<delta_on` or `abs(heff)<delta_h`; if the previous state was fine, retain it until `abs(Keff-Kc)>delta_off`; route fine patches through the expensive block and coarse patches through pooling plus a cheap block. 3. The paper supplies the effective-parameter construction and the critical coupling; \(\delta_K\), \(\delta_h\), the pooling operator, and residual neural blocks are estimated by validation. 4. First cheap experiment: use CIFAR-10 with synthetic decorated image patches, comparing fixed-resolution hierarchical processing against the controller at matched accuracy and FLOPs. Estimate feature correlation length from spatial autocorrelation. The predicted signature is \(\xi\) increasing approximately as \(|K_{\mathrm{eff}}-K_c|^{-1}\), with the controller retaining the most tokens in a band around \(K_c\). Under a sweep of the synthetic coupling, the mean token count should switch near the prescribed hysteresis thresholds rather than at an arbitrary training-dependent point.

## Disclaimer

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