Highly successful learning based on modified Catmull-Rom spline activation function

A sigmoidal-like activation function, a modified Catmull-Rom spline activation function, is proposed. The advantages of the proposed activation are low complexity of the hardware implementation and arithmetic operation. Since it gives a good error signal for updating the weight values, the usual gradient based training algorithm can be used to train the network with the proposed activation function. Our experiments based on the well established benchmarks, such as LT, SONAR, IRIS, OCTANE, and 5-BIT COUNT data sets, show that the convergent speed and the successful rate based on the proposed model is more than those of the existing models using the sigmoidal function. The network obtained by using the proposed activation function is more compacted than the network obtained under the sigmoidal function in term of the arithmetic operations and the network components