Symmetry in the Algebra of Learning: Dual Numbers and the Jacobian in K-Nets

The black-box nature of deep machine learning hinders the extraction of knowledge in science. To address this issue, a proposal for a neural network (k-net) based on the Kolmogorov–Arnold Representation Theorem is presented, pursuing to be an alternative to the traditional Multilayer Perceptron. In its core, the algorithmic nature of neural networks lies in the fundamental symmetry between forward-mode and reverse-mode accumulation techniques, both of which rely on the chain rule of partial derivatives. These methods are essential for computing gradients of functions, an operation that is at the core of the training process of neural networks. Automatic differentiation addresses the need for accurate and efficient calculation of derivative values in scientific computing; procedural programs are thus transformed into the computation of the required derivatives at the same numerical arguments. This work formalizes the algebraic structure of neural network computations by framing the training process within the domain of hyperdual numbers. Specifically, it defines a Kolmogorov–Arnold-inspired neural network (k-net) using dual numbers by extending the univariate functions and their compositions that appear in the representation theorem. This approach focuses on computation of the Jacobian and the ability to implement such procedures algorithmically, without sacrificing accuracy and mathematical rigor, while exploiting the inherent symmetry of the dual number formalism.