NN1: Baseline Subtraction: From the Vacuum to the Weight Matrix — A Regularization Method Derived from the Cosmological Constant Problem

The cosmological constant problem is the largest discrepancy between theory and observation in physics. Quantum field theory predicts a vacuum energy density 10^{120} times larger than observed. The Canvas Model resolves this through baseline subtraction: spatially uniform field configurations produce no observable effects. Only spatial variations gravitate. The enormous vacuum energy does not gravitate because it is uniform. The cosmological constant is small because the uniform part does not contribute. This paper shows that the same principle improves neural network training. The symmetric part of a weight matrix—the baseline that responds identically to all inputs—carries no discriminative information. Subtracting it during training reduces overfitting without reducing capacity. We call this baseline subtraction regularization. What this paper provides: · A derivation from first principles. The method is not a heuristic. It follows from the Threshold Condition of the Canvas Model—the same condition that prevents the vacuum from destroying the universe. The symmetric part of a weight matrix is the S-invariant baseline; the antisymmetric part encodes directional relationships. Only the antisymmetric part should be trained.· An operational procedure for any feedforward network. For square weight matrices, compute W_{\text{train}} = \frac{1}{2}(W - W^\top) and train only this component. The symmetric baseline never contributes to the network's output, removing a source of overfitting.· Generalization to rectangular matrices and convolutional layers. Two approaches are provided: block decomposition (simple, works when dimensions are within a factor of 2) and canonical decomposition via SVD (more principled, handles any aspect ratio). For convolutional kernels, the symmetry operator acts by rotating the kernel 180 degrees and flipping the sign.· Empirical validation on MNIST and CIFAR-10. On MNIST, baseline subtraction alone achieves 98.42% test accuracy (vs 98.21% for standard training) and reduces the overfitting gap by more than half. On CIFAR-10, baseline subtraction alone achieves 89.23% test accuracy (vs 88.12% for standard training) and reduces the overfitting gap by 42%. The method is complementary to dropout; the combination achieves the best results on both benchmarks. Training time overhead is approximately 2%.· Comparison to existing methods. Baseline subtraction is not dropout, weight decay, or batch normalization. Those methods were discovered empirically; baseline subtraction is derived from a physical law. It operates on a different aspect of the network and is complementary to existing regularization techniques. Why this matters: The principle is not a heuristic. It is a consequence of the same physical law that prevents the vacuum from destroying the universe. The symmetric weight matrix is the vacuum—present, but inert. The antisymmetric part is the structure. Only structure matters. The vacuum does not gravitate. The symmetric weight matrix does not discriminate. The principle is the same. The mathematics is the same. The result is the same: subtract the baseline, train the deviations, and watch the structure emerge. Keywords: baseline subtraction, cosmological constant problem, neural network regularization, overfitting, symmetric weight matrix, antisymmetric weight matrix, threshold condition, Canvas Model, MNIST, CIFAR-10, dropout, weight decay