Pinning the Glass: Diagnosing and Breaking Kinetic Arrest in Oscillator-Based Feature Binding

Some neural networks solve the "binding problem" (keeping track of which color goes with which shape when many objects share one image) through timing rather than dedicated memory. Each unit carries a phase, like the position of a clock hand. Units describing the same object rotate into step with one another, and units describing different objects drift apart. To produce a grouping, such a network relaxes iteratively, and it does so inside a fixed budget of update steps. This proposal claims that when these networks fail on cluttered, textured scenes, the cause is often not a lack of representational capacity. They get stuck. The phase configuration freezes into a wrong but stable grouping before it finds the correct one. Glass-forming liquids show the same behavior: as they cool, their rearrangements slow down faster than any simple energy barrier would predict.

If the analogy holds, glass physics supplies both a measurement and a fix. The measurement is fragility, which describes how sharply relaxation time blows up as the system is made "colder." The fix is random pinning: freezing a random fraction of particles in a typical configuration is known to bring a liquid into its glass state without the astronomically slow cooling otherwise required. We propose to measure fragility in a Kuramoto-oscillator network on textured multi-object scenes, test whether it predicts binding failures image by image, and test whether pinning a small random fraction of oscillators to phases sampled from the network's own short runs beats the obvious alternatives at matched compute. The obvious alternatives are more iterations, noise annealing, best-of-K restarts, and spatial phase seeding.

Core Claim

Phase-based binding models are now competitive. Artificial Kuramoto Oscillatory Neurons (AKOrN) replaces standard neurons with coupled unit-vector oscillators. It reports competing with slot models on ClevrTex and reaching 89.5% on out-of-distribution Sudoku versus 30.3% for a recurrent transformer. The authors note that lower energy tends to accompany correct solutions. KoPE adds Kuramoto phases to ViT attention and gains 2.8 mIoU on ADE20K. Both papers report sensitivity to initialization and the cost of iterating. KoPE learns much more slowly from random phases than from phases seeded with 2D coordinates, and CV-RNN needs first-frame phase seeding. None of these works asks what kind of landscape the iterations are traversing.

Our claim has three parts, each falsifiable. First, the oscillator network's relaxation time grows super-Arrhenius as injected phase noise is lowered. Super-Arrhenius means the effective barrier itself grows, rather than staying constant. This growth should be markedly steeper on textured and camouflaged scenes than on plain ones. Second, a per-image fragility estimate predicts segmentation failure better than generic difficulty statistics such as object count or final energy. Third, random pinning beats matched-compute annealing and restarts. It should do so specifically as a threshold in the pinned fraction, not as a smooth gain.

The third prediction is what separates pinning from annealing. Glass theory says slow cooling of a fragile system saturates, because the required time diverges (Cavagna; Cammarota et al.). Pinning instead removes configurational freedom, the number of available metastable states, without supercooling (Ideal glass transitions by random pinning; Random pinning glass transition: hallmarks, mean-field theory and RG).

Research Program

Defining the measurement. The base model is AKOrN trained unsupervised on ClevrTex as in its paper. At test time we inject isotropic noise of strength $T$ into each oscillator update and treat $T$ as a temperature. For each $T$ we run the dynamics from many random initial phases. We track the overlap $q(t)$, defined as the agreement of the token-by-token grouping (pairwise same-cluster indicators) at step $t$ with the grouping at a late reference step. The relaxation time $\tau(T)$ is the number of steps for $q$ to reach its plateau within a tolerance.

We fit two models. The Arrhenius model is $\log\tau = A + E/T$. The Vogel–Fulcher form is

$$\log \tau(T) = A + \frac{B}{T - T_0},$$

where $A$, $B$ and $T_0$ are fitted constants and $T_0>0$ marks apparent divergence. We define the network's glass temperature $T_g$ as the noise level at which $\tau$ equals its iteration budget, so the definition is tied to the deployed compute. The fragility index is $m = d\log\tau / d(T_g/T)$ evaluated at $T=T_g$, the standard steepness measure transplanted. A strong system has a small $m$ and an Arrhenius fit. A fragile system has a large $m$ and a clearly better Vogel–Fulcher fit. We compute $m$ per dataset and, using fewer replicas, per image.

Dynamical heterogeneity. We also measure which tokens relax slowly. The analogue is the four-point susceptibility $\chi_4$ studied in elasto-plastic scaling models, here taken as the variance across replicas of $q(t)$. We then check whether slow tokens cluster on texture-camouflage boundaries, where ClevrTex documents slot models fracturing into tiles. Those elasto-plastic models are Arrhenius by construction, so they serve only as a template for the measurement and not as a predictor of fragility.

The intervention. At test time we run one short replica for $k$ steps from random phases. We then select a uniformly random fraction $c$ of oscillators and clamp each to its phase in that replica. The remaining oscillators relax under the full dynamics for the rest of the budget. Physics pins particles to an equilibrium configuration. We have no ground-truth equilibrium, so the short replica stands in for a typical sample, and $k$ is a controlled variable. If pinning after very short runs works, the benefit comes from breaking symmetry. If it needs long runs, the benefit comes from inheriting good structure.

A training-time variant applies the same pinning stochastically during unsupervised training. It is a cheap ingredient that addresses the initialization sensitivity reported for KoPE and CV-RNN without coordinate seeding or mask supervision. A further test checks whether the gain generalizes to tracking. On FeatureTracker we pin a random fraction of oscillators to the previous frame's phases rather than seeding the whole frame, as CV-RNN does.

A contradiction we address directly. Sengupta et al. and the random-pinning literature report that the sharp pinning transition is suppressed in two dimensions. Image tokens live on a 2D grid, so this should worry us. We turn it into a prediction instead. Where coupling is local, as in convolutional coupling, the pinning gain should be smooth and weak. Where coupling is non-local through attention, as in AKOrN's attentive variant, the effective geometry is closer to mean-field and a threshold in $c$ should appear. If both regimes look the same, the physical reading is wrong.

Validity constraint. The landscape interpretation requires an energy that the dynamics actually descend, which needs approximately symmetric coupling. We will report the asymmetry of the learned couplings. Where asymmetry is large, we will treat the fragility index as a purely dynamical, operational statistic.

Baselines

The naive reference is AKOrN run with its default iteration count and random initialization. We retrain it ourselves on ClevrTex with the released recipe. The file reports no absolute AKOrN ClevrTex number, only the claim that it competes with slot models, so we found no reported result on Lacuna for AKOrN mIoU on ClevrTex and will establish it ourselves. For context, the ClevrTex paper reports mIoU of around 42% for GNM, 29% for IODINE and 0% for SPAIR on ClevrTex, against about 66% for SPAIR on plain CLEVR. These anchor what "competitive" means, but they are not our comparison.

The strongest competitor is the same trained AKOrN given the same total number of oscillator updates as the pinning procedure, spent in each of the ways a practitioner would actually try. The first uses more iterations. The second uses a tuned noise-annealing schedule, with initial temperature, schedule shape and length tuned on a validation split using the same number of tuning trials as pinning's $c$ and $k$. The third runs best-of-K random restarts and keeps the lowest-energy result, with K chosen so that total steps match. The fourth applies KoPE-style 2D-coordinate phase seeding to all oscillators. Restarts and seeding are the strongest threats. Restarts already exploit the multi-replica structure we rely on, and seeding is the documented fix for slow symmetry breaking.

The ablation that isolates the key ingredient pins a random fraction to random phases instead of phases from the short replica. It is crossed with confidence-selected pinning, which pins oscillators whose grouping agrees across replicas instead of a random subset. Random-phase pinning tests whether any symmetry breaking suffices. Confidence selection tests whether randomness of the pinned set matters, as theory says it should, for avoiding biased pinning. Pinning counts as beating the field only if it improves ClevrTex foreground mIoU over the best compute-matched baseline by more than three seeds' spread, both in-distribution and on the camouflage and out-of-distribution splits.

Why not just…?

Why not just run more iterations? If failures were simply slow convergence along an Arrhenius-like path, extra steps would help steadily and pinning would add nothing. The fragility measurement settles this directly. If ClevrTex relaxation fits Arrhenius with a low $m$, we expect added iterations to match pinning, and we would report that the glass framing is unnecessary for this model. Our hypothesis predicts the opposite: in a fragile regime, gains from extra steps flatten well before the error floor.

Why not just anneal the noise? Annealing is the textbook response to a rough landscape and is a genuine threat. The distinguishing evidence is the functional form of the gains. Annealing gains should rise smoothly with schedule length and saturate where $\tau(T)$ diverges. Pinning gains should jump near a critical fraction under non-local coupling. If tuned annealing matches pinning at equal compute on the camouflage split, pinning is redundant, and we will say so. The diagnostic contribution would survive that outcome.

Why not just seed phases with spatial coordinates, or use slots? KoPE shows that coordinate seeding speeds learning. However, seeding imposes a spatial prior on every token, which is exactly the prior that produces tiling on textured scenes. Slot models that bake in object allocation collapse on ClevrTex (ClevrTex; compare Neural-EM). We include seeding as a baseline. If it wins in-distribution but loses on the camouflage split, that matches our account. If it wins everywhere, our account fails.

Proposed Experiments and Decisive Checks

The first check is the fragility comparison. We expect a Vogel–Fulcher fit significantly better than Arrhenius, with a larger $m$ on ClevrTex than on CLEVR, and $m$ increasing with object count and in the camouflage split. The second is prediction. Per-image $\tau$ at the deployed noise level should predict per-image mIoU failure, measured by AUROC for flagging the worst quartile, beyond object count and final energy combined. The third is the pinning sweep over $c\in[0,0.5]$ and $k$, run under both local and attention-based coupling, looking for a threshold only in the latter.

The fourth links back to binding as defined in Formalizing the Binding Problem, which estimates binding information with probes and reports that ViT spatial tokens retain 92.2–97% of it on ColorShape while the summary token resolves only 48.5%. We apply its deep-network probe to oscillator states on ColorShape before and after pinning. The question is whether pinning increases decodable conjunction information or only sharpens segmentation boundaries; that paper notes decodability and use can diverge.

The fifth is a transfer check of the training-time variant on FeatureTracker. We compare against CV-RNN's reported ~90% accuracy, which was achieved with a supervised synchrony loss. A pinned unsupervised model approaching it would address the open question of unsupervised synchrony, and falling well short would bound the claim.

The whole program fails cleanly under either of two outcomes: Arrhenius relaxation on textured scenes, or no threshold under non-local coupling. Either result would itself be informative for anyone extending KomplexNet-style oscillator layers to natural images.