Dynamics of a Discrete Memristor-Based Adaptive Activation Gradient Heterogeneous Neural Network

The activation function is an indispensable component of neural networks, and its activation gradient reflects the response speed of neuronal electrical activity. Neural networks constructed from neurons with different activation functions are more consistent with the biological plausibility of neural networks. This paper proposes a Discrete Memristor-based Adaptive Activation Gradient Heterogeneous Neural Network (DMAAGHNN), in which a discrete memristor is introduced into the activation functions of neurons to serve as adaptive activation gradients. The proposed network is composed of three different types of activation functions. The influences of coupling strength, neural decay coefficient, and weight on the system’s dynamics are studied, revealing chaos, hyperchaos, multiple firing patterns, coexistence of heterogeneous chaotic attractors, and high spectral entropy complexity. Furthermore, an Field Programmable Gate Array (FPGA)-based hardware circuit of the DMAAGHNN is successfully implemented. The chaotic sequences generated by the system are applied to a pseudorandom number generator. All generated sequences pass the National Institute of Standards and Technology (NIST) test suite, indicating that the system possesses good randomness.