Memory level neural network: A time-varying neural network for memory input processing.

Affective computing is an important foundation for implementing brain-like computing and advanced machine intelligence. However, the instantaneous and memory fusion input characteristic makes current neural networks not suitable for affective computing. In this paper, we propose an affective computing oriented memory level neural network. A "switch" has been added to the memory level neurons, which will achieve a transition from the instantaneous input to the memory input when the temporal integration of inputs above a certain threshold. Then, the "switch" is continualized by an adjustable sigmoid function whose parameters are tuned to adjust the speed of the transition and the mixing ratio of the two inputs. Multiple memory level neurons form a deep time-varying neural network capable of handling fusional inputs. We demonstrate on both process datasets and static datasets that the memory level neural network successfully converges on both datasets and meets the error accuracy requirements. [ABSTRACT FROM AUTHOR]