Towards Artificial Hydrocarbon Networks: The Chemical Nature of Data-Driven Approaches

Challenges and Issues on Artificial Hydrocarbon Networks: The Chemical Nature of Data-Driven Approaches

“…To this end, the mixture constitutes a model [1]. Thus, the inspiration in organic compounds to develop a machine learning method considers three facts observed from nature [3]: (i) stability as the property of compounds to maintain their geometric configurations; (ii) organization based on the ground-state principle aiming to preserve energy minimization within the compounds; and (iii) multi-functionality for promoting transfer learning. For training purposes, the method considers the simple AHN training algorithm [1] which is based on the gradient descent and the numerical solution of least squares estimates via QR-factorization.…”

“…This training algorithm has reported well performance in predictive power for low-dimensional input spaces, and large training time for computing suitable parameters in the model [4]. Currently, new training methods have been proposed based on hierarchical training [3] or using stochastic-parallel metaheuristic optimization [4]. The later, accelerating training in more than 3,500 times the simple AHN training algorithm.…”

“…Literature reports many different applications of AHN. Those can be classified as follows: function approximation and modeling [1]; robust human activity recognition systems [5]; signal processing in denoising audio and face recognition [1,6]; online advertising [6]; intelligent control systems for robotics [1,7,8,11] and mechatronics [1,9,10]; bio/medical applications [5,6,12]; and, theoretical approaches such as hybrid fuzzy-molecular inference systems [8], interpretability of the model [12] and training algorithms [3,4].…”