Abstract: Recent deep learning models increasingly rely on depth without structural guarantees on the validity of intermediate representations, rendering early stopping and adaptive computation ill-posed. We address this limitation by formulating a structural requirement for state-space model's scale-consistent latent dynamics across iterative refinement, and derive Fractal of Stationary Transformations (FROST), which enforces a self-similar representation manifold through a fractal inductive bias. Under this geometry, intermediate states correspond to different resolutions of a shared representation, and we provide a geometric analysis establishing contraction and stable convergence across iterations. As a consequence of this scale-consistent structure, halting naturally admits a ranking-based formulation driven by intrinsic feature quality rather than extrinsic objectives. Controlled experiments on ImageNet-100 empirically verify the predicted scale-consistent behavior, showing that adaptive efficiency emerges from the aligned latent geometry.
Introduction
In modern deep learning, the standard paradigm is that "deeper is better." However, there is rarely a structural guarantee that the representation at an early layer resembles the final output. In most models, intermediate layers act as "incomplete artifacts"—precursors that lack intrinsic utility on their own. This creates a significant challenge for adaptive computation (early exiting). If one stops a model early to save resources, they typically rely on auxiliary halting modules or losses that often conflict with the main task, leading to unstable training or "representation drift."
FROST (Fractal of Stationary Transformations) addresses this by treating depth not just as a sequence of layers, but as a process of iterative refinement. By injecting a fractal inductive bias into the architecture, FROST ensures that intermediate representations are geometrically consistent with the final output—effectively turning depth into a "zoom" lens for data.
Methodology
FROST reformulates the latent dynamics of a State-Space Model (SSM) using fractal geometry to ensure that as the model iterates, it refines the same representation at a higher resolution rather than moving to a different distribution.
- Fractal Inductive Bias & Stationary Transformations: The model utilizes weight-tied, stationary transformations to ensure the rules of the geometry remain constant across depth. This forces the latent trajectory to be self-similar.
- Spectral Gating ($\lambda$): A learnable scaling factor, $\lambda$, acts as a "resolution parameter." It controls the contraction rate of the model, ensuring it converges stably to a fixed point and remains resistant to exploding gradients.
- Ranking-based Halting: Because representations are scale-consistent, the model can use a simple ranking system to decide when to stop. It compares the "difficulty" of a sample relative to others in a batch using a streaming quantile sketch. If a sample's features are already high-quality, the model halts, disentangling the halting decision from the learning objective.
Empirical Results
The framework was evaluated on ImageNet-100 and CIFAR-100 using ResNet and Vision Transformer (ViT) backbones.
Structural Consistency
Mathematical proofs and empirical verification show that FROST's intermediate states have much higher cosine similarity to the final output than vanilla models. As shown in the table below, FROST achieves alignment scores near 0.90 by iteration 6, whereas standard SSMs struggle with unstable, near-orthogonal representations (often around 0.05 in early stages).
Efficiency and Information Density
On a ViT-B/16 backbone, FROST reduced the average refinement depth from 12.0 to 4.53 iterations. This resulted in a throughput jump from 21.4 to 42.2 images per second—nearly a 2x speedup with negligible loss in accuracy. Furthermore, Minkowski dimension analysis indicates that FROST learns a "space-filling" trajectory with higher dimensionality than baselines, suggesting it packs more information into the same number of parameters.
Conclusion
FROST provides a structural solution to the problem of representation drift in deep models. By reinterpreting depth as resolution through a fractal inductive bias, it enables reliable early halting based on intrinsic feature quality. While future work may explore token-level halting and larger-scale validation (e.g., ImageNet-1k or LLMs), FROST establishes a robust geometric foundation for scale-consistent state-space dynamics.
Summary updated