Recurrent Neural Networks for Computing Pseudoinverses of Rank-Deficient Matrices
SIAM Journal on Scientific Computing
IEEE Transactions on Computers
| Hi-index | 0.01 |
In this paper, we show how the diffusion of the Nitric Oxide retrograde neuromessenger (NO) in the neural tissue produces Diffusive Hybrid Neuromodulation (DHN), as well as positively influencing the learning process in the artificial and biological neural networks. It also considers whether the DHN, together with the correlational character that helps Hebb's law, is the best schema to represent the resulting learning process. We conducted the entire study by means of analysing the behaviour of the Diffusion Associative Network (DAN) proposed by our group [1]. In addition, and in order to identify the way in which the diffusion of the NO may coincidently affect the learning process, such as those supported by Hebb's Law, we created and studied recursive schemas of the calculation of the optimal weight matrix in the lineal association, which were then used to try to identify possible ways to express the diffusion. From this last study, we concluded that the recursive schemas are not sufficient in the calculation of the weight matrix in order for the expression identification of the diffusion phenomena in the learning process. These results pointed us towards the search for new schemas for the Hebb law, based on the effect of the NO and that it is different to the modulated correlation, as well as for the search for a more appropriate diffusion model for the NO, such as the compartimental model.