NEURON MODEL SIGMOID ACTIVATION FUNCTION BASED ON MULTI-OPTIONAL FUNCTIONS ENTROPY CONDITIONAL OPTIMIZATION DOCTRINE
Journal Title: Електроніка та системи управління - Year 2018, Vol 4, Issue 58
Abstract
It is made an attempt to discover an explainable plausible reason for a neuron activation function, of a sigmoid type function like logistic function, substantiation in terms of the multi-optional conditional optimality doctrine for the special hybrid-optional effectiveness functions uncertainty. In the studied case, the input-output mapping is stipulated by the entropy of the activation function conditional optimal distribution in regards with the induced local field of the neuron. It is proposed to evaluate the direction of uncertainty with the combined hybrid relative pseudo-entropy function. This is a new insight into the scientific substantiation of the well-known dependency derived in another way. The developed theoretical contemplations and mathematical derivations are verified with numerical simulation and plotted diagrams.
Authors and Affiliations
Andriy Goncharenko
Keywords
Related Articles
- EP ID EP514740
- DOI 10.18372/1990-5548.58.13518
- Views 173
- Downloads 0
ALGORITHM OF VARIATIVE FEATURE DETECTION AND PREDICTION IN CONTEXT-DEPENDENT RECOGNITION
Application of context-dependent classification for recognition tasks is proposed. In the context-free classification, the starting point was the Bayesian classifier. Morphological features such as object form, area, and...
SYSTEM OF ADAPTIVE TASKS SELECTION IN THE TESTING PROCESS
In this article, an automated knowledge testing system with adaptive parameters, which assumes the individualization of the training courses and test tasks content designed to control the knowledge of each learner is con...
SAFETY ENHANCEMENT ALGORITHMS OF THE MICRO-UAV PILOTING
The article is devoted to algorithms development of piloting with enhanced safety for micro-UAV. The critical situations arising during the flight, typical failure of the electrical equipment of the ground station commun...
ALGORITHM OF PRUNING OF HYBRID NEURAL NETWORKS ENSEMBLE
Despite the fact that the ensemble is usually more accurate than a single network, existing ensemble techniques tend to create unreasonably large ensembles that increase the use of memory and computation costs. The ensem...
FEATURES OF MULTIPLICATIVE COMPLEMENTARY SIGNAL CONSTRUCTIONS BASED ON GENERALIZED BINARY BARKER SEQUENCES
Radio and signal processing systems often use spread-spectrum and pulse compression techniques combined with the matched filtering approach as the main signal processing scheme. Among different signals that have been des...