NN2: The S-Invariant Attractor: A Principled Criterion for Neural Network Convergence

Neural network training stops when the validation loss plateaus—"early stopping." The threshold is arbitrary: if the loss hasn't improved in k epochs, stop. This can halt training before convergence is complete or continue long after overfitting has begun. The choice of k is a hyperparameter with no theoretical grounding. This paper proposes a principled convergence criterion motivated by the Canvas Model: training is complete when the weight matrix reaches its S-invariant attractor—the configuration where each degree of freedom is individually invariant under the symmetry operator \mathcal{S} that exchanges internal and external primitives. For a weight matrix W, motivated by the primitive structure of the Canvas Model, we define \mathcal{S}[W] = -W^\top. This choice reflects the exchange of internal (property) and external (dynamic) primitives, with the minus sign arising from the Polarity primitive (P4). The S-invariant attractor is W = -W^\top: an antisymmetric matrix. Convergence is not "the loss stopped decreasing." Convergence is W_{\text{base}} \to 0, where W_{\text{base}} = \frac{1}{2}(W + W^\top) is the symmetric baseline. The paper defines the convergence measure C(W) = \|W_{\text{base}}\|_F / \|W\|_F and validates it on MNIST and CIFAR-10. The measure correlates with validation loss (r = 0.84 [95% CI: 0.77, 0.89] on MNIST; r = 0.81 [95% CI: 0.70, 0.88] on CIFAR-10). Stopping at C(W) < 0.01 achieves accuracy comparable to or exceeding early stopping, with substantially lower sensitivity to arbitrary parameters. The paper compares against alternative symmetry operators and standard convergence criteria (gradient norm, weight change, loss change). The measure also provides early warning of overfitting and can diagnose architectural inadequacy. Why this matters: The S-invariant attractor concept is universal in the Canvas Model: it governs the Riemann zeros settling onto the critical line, the cosmological constant drifting toward its observed value, and any gradient-driven system evolving toward equilibrium. The specific realization \mathcal{S}[W] = -W^\top for neural networks is a motivated instantiation of this principle. The empirical results provide the primary justification for this choice. The criterion is simple (C(W) = \|W_{\text{base}}\|_F / \|W\|_F, stop when C(W) < 0.01), robust to its parameter, and provides diagnostic capabilities beyond early stopping. It eliminates the arbitrary patience hyperparameter, provides early warning of overfitting, and can diagnose architectural inadequacy. Keywords: neural networks, convergence criterion, early stopping, S-invariant attractor, Canvas Model, symmetry operator, generalization, deep learning