Biological neural networks are based on the parallel architecture of animal brains. The brain is a mesh of billions and billions of neurons. The process of neural communication, that gives the brain most of its functional complexity, has been a key area of research for many years now. There has been a great deal of effort to understand the processes of information storage and retrieval going on inside the brain. “100-step limitation”.
Artificial neural networks (ANN) are a form of multiprocessor computer system with:
simple processing elements (caled neurons)
a high degree of interconnection (connections and link weights)
simple scalar messages (summation and multiplication)
adaptive interaction between elements (data transfer between neurons)
Simon Haykin provides a general definition of neural learning:
Learning is a process by which the free parameters of a neural network are adapted through a continuing process of stimulation by the environment in which the network is embedded. The type of learning is determined by the manner in which the parameter changes take place.
In artificial neural networks, learning refers to the method modifying the weights (as well as topology if the network is self-organizing) of connections between the nodes in a specified network. The rule followed to update the connection weights – the learning rule – determines how well the network converges towards its desired optimality.
CWT's ANN research includes the following topics:
one-class-one-network (OCON) ANN topology design for on-line DSP implementation [Le 2005]
advanced learning algorithm (modified BP or pdf-based training) design for adaptive feature recognition
feature space definition and optimization for radio domain information handling
complex-valued network implementation for waveform features
activation function design for the trade-off between robustness and accuracy in performance