No abstract

This paper presents an improved extension of the previous algorithm of the authors called KAdam that was proposed as a combination of a first-order gradient-based optimizer of stochastic functions, known as the Adam algorithm and the Kalman filter. In the extension presented here, it is proposed to filter each parameter of the objective function using a 1-D Kalman filter; this allows us to switch from matrix and vector calculations to scalar operations. Moreover, it is reduced the impact of the measurement noise factor from the Kalman filter by using an exponential decay in function of the number of epochs for the training. Therefore in this paper, is introduced our proposed method sKAdam, a straightforward improvement over the original algorithm. This extension of KAdam presents a reduced execution time, a reduced computational complexity, and better accuracy as well as keep the properties from Adam of being well suited for problems with large datasets and/or parameters, non-stationary objectives, noisy and/or sparse gradients.