Advanced Soft Computing Technique: Introduction of Thousands of Non-linear Specific Activation Functions into the Deep Learning Artificial Neural Networks

Non-linear activation functions play an extremely crucial role in neural networks by introducing non-linearity. This nonlinearity allows neural networks to develop complex representations and functions based on the inputs that would not be possible with a simple linear function. Without a non-linear activation function in the network, a neural network, no matter how many layers it had, would behave just like a single-layer perceptron. So, why is increasing non-linear specific activation functions desired? What effect do they have on the overall performance of the network? The author applied the modified neural network architectures proposed in Jamilu Adamu (2019) to linked the most volatile Chicago City daily maximum temperature data exported from NASA POWER| Data Access from 31st March, 2021 to 31st March, 1981 (roughly 40 years). With the help of Jameel’s ANNAF Deterministic and Stochastic Criterion, 86 specific activation functions were generated using Superneunet software prototype developed by the author. With the help of QI Macros, Normality Test confirmed Chicago maximum temperature data has being Non-normal, volatile and unpredictable. To assess the performances between the trial and error traditional activation functions with that of modified, Actual and Initial Predicted values of Chicago City maximum temperature were correlated, Tanh (0.947) outperformed ReLu (0.945) and Sigmoid (0.940) using traditional trial and error activation functions. While, with the modified specific activation functions, Cubic-Exponential-Gaussian-Gaussian (0.928) outperformed Exponential- Exponential-Linear (0.910) and Cubic-Exponential (0.923). Note that with minimum number of layers and nodes, the activation functions were selected arbitrarily and kept constant throughout the hidden layers of Superneunet prototype. With automation, the modified functions can outperform the traditional out rightly. The results of the study suggests that the technique can be used to specifically, accurately and precisely predict any time series application and the prices of over 35 million instruments across all asset classes traded 24/7 in 160 countries aggregated from 330 exchanges of Bloomberg Terminal and other world’s largest financial and weather assets terminals as against the current practice that uniformly uses traditional trial and error few activation functions across all areas of application. Now the current version of Superneunet software prototype accommodated up to 100+ specific activation functions and can be extended to 500 or even thousands activation functions per area of application. So, what will happen to our predictions or forecasts if we extend Superneunet to accommodate millions or billions of specific activation functions? Do we need to upgrade Superneunet to accommodate this Quantum Computers application? One of the fundamental benefits of this novel breakthrough, it can waive years of research by researchers around the globe for constructing single desirable activation function, with it can be done by just clicking of a button. This can completely eliminate phobia and psychological trauma experiences when selecting the right activation functions, thus can fantastically save cost, time, energy, simultaneously accelerate AI technology production and global economic growth. To crown it up, the humanity can benefit from the results of this study through Environment, Finance, Education, Health & Pharmaceuticals, Transportation & Security, Science & Technology, and Innovation & Creativity