IDEA #8ASPVW JND and Shape Dependent Neural Network with Trainable Activation Function. 21A0026

A method and system disclose for an artificial neuron, which is the main block of any neural network system. The neuron(s) may (1) generate by at least two nonlinear combinations of multiple inputs, with various biases, (2) identify the insignificant parameters or neurons and remove them from the neural network, and (3) update (adaptively, not adaptively, or dynamically) or train artificial neurons from application needs. Moreover, a method and system for extra disclose including, (4) the innovative activation functions; (5) the input data de-correlated both integer and non-integer transform-based neural networks (6) the convolution parameters reduced shape-dependent filters, and (7) the transform-filters coefficients based JND defined information furthered neural networks, (8) parameter/operation guided and adaptive frequency domain CNN architecture, and (9) the JND-ed trainable transform domain neural networks architecture. The disclosed filters can operate with a commonly used convolutional neural network. The use of the innovative main block may allow for improved efficiency of a deep learning system, reduce simulation time, decrease the number of operations, and produce a trimmed neural network. The innovative artificial neurons have great potential for neuromorphic computing and various real-life applications.