# Distribution-Preserving Fragmentation Augmentation

- ID: 414
- Canonical URL: https://synthcore.org/idea/414/distribution-preserving-fragmentation-augmentation
- API JSON: https://synthcore.org/api/idea/414.json
- API Markdown: https://synthcore.org/api/idea/414.md
- Verification status: unverified
- Source: [arXiv:2607.04807](https://arxiv.org/abs/2607.04807)
- Category: regularization
- Solves: generalization, accuracy, stability
- ML areas: cnn, graph-nn, training, data-augmentation
- Math tags: pde, dynamical-systems, geometry
- Ratings: usefulness 4/10; difficulty 4/10; novelty 7/10

## Idea description

Augment spatial training examples by replacing a compact active region with several separated components while preserving its exact value histogram, total active area, and amplitude. The augmentation probes the nonlinear interaction between diffusion-like receptive fields and threshold activations, which the paper shows can make fragmented and compact inputs evolve in opposite directions despite identical distributions.

## Mathematical statement

The paper studies the reaction-diffusion dynamics ∂_t u = ∂_xx u + f(u) for bounded nonnegative u, with a bistable nonlinearity satisfying f(0) = f(θ) = f(1) = 0, f'(0) < 0, f(s) < 0 for 0 < s < θ, f(s) > 0 for θ < s < 1, integral from 0 to 1 of f(s) ds greater than 0, and f(s) < 0 for s > 1. Here θ is the activation threshold. Theorem 1.6 states that when 0 < R_1 < L_infinity < R_2, there exist an amplitude α > θ and an integer n ≥ 2 such that the compact datum u_0^blk(x) = α times the indicator of (-R_1/2, R_1/2) extinguishes, while the fragmented datum u_0^frag(x) = α times the indicator of E_n(R_1,R_2) propagates to 1; both have the same distribution function and active measure R_1. Define E_n(R_1,R_2) as n equal intervals, each of length R_1/n, uniformly spaced inside an envelope of length R_2, so its measure is R_1. The transferable property is that threshold dynamics depend on spatial connectivity and gap size, not only on pointwise statistics.

## Key formulas

- $$\partial_t u=\partial_{xx}u+f(u),\qquad u(0,x)=u_0(x)$$
- $$f(0)=f(\theta)=f(1)=0,\quad f'(0)<0,\quad f(s)<0\ (0<s<\theta),\quad f(s)>0\ (\theta<s<1),\quad \int_0^1 f(s)\,ds>0$$
- $$u_0^{\rm blk}(x)=\alpha\mathbf{1}_{(-R_1/2,R_1/2)}(x),\qquad u_0^{\rm frag}(x)=\alpha\mathbf{1}_{E_n(R_1,R_2)}(x),\qquad |E_n(R_1,R_2)|=R_1$$
- $$\mathcal{L}(u_0^{\rm blk})=\mathcal{L}(u_0^{\rm frag})\quad\text{but}\quad \lim_{t\to\infty}u^{\rm blk}(t,\cdot)=0,\qquad \lim_{t\to\infty}u^{\rm frag}(t,\cdot)=1$$

## Implementation notes

Integrate this at the input or intermediate-feature augmentation stage of a CNN, U-Net, neural cellular automaton, or spatial graph network. For a tensor X with shape C by H by W, select one channel or a learned scalar mask M. Threshold it into an active set, preserve the active values, and replace the active set by n separated components with the same total pixel count. Choose n from {2, 4, 8}, place equal-area components inside an envelope expanded by a factor gamma, and copy the original active amplitude α. Apply the transformation with probability p while keeping the original label. Pseudocode: mask = (X[c] > tau); A = mask.sum(); n = sample({2,4,8}); E = place_equal_components(area=A, envelope=gamma * bounding_box(mask), n=n); X_aug[c] = alpha * E + X[c] * (1-mask). The preserved mathematical quantities are the value distribution, active measure R_1, fragmentation count n, and envelope scale R_2. Estimate tau from feature quantiles and choose gamma and p on a validation split. First test on MNIST or CIFAR-10 with a small CNN and synthetic objects whose connectedness varies. Compare ordinary augmentation, histogram-preserving pixel shuffling, and structured fragmentation at equal compute. Measure clean accuracy, accuracy under object splitting, calibration, and performance as n increases. Success means improved accuracy on fragmented layouts without degrading compact-layout accuracy, together with lower prediction variance across spatial arrangements having the same activation histogram.

## Disclaimer

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