Noise Profiling for ANNs: A Bio-inspired Approach

Artificial neural networks (ANNs) are potent computational models, which are capable of completing a range of perception-related tasks. However, sometimes it is difficult for them to learn from complex data. Therefore, it is preferable to introduce noise into the input or hidden layers of the ANN during model training in order to get around this problem. As a result, it can enhance the adaptability of the model. This paper is an approach to noise profiling for ANNs that draws inspiration from the biological workings of insect sensory systems. By using specialized sense organs, insects are evolved to deal with noisy environments. The using of both Gaussian and Chaotic noises have various statistical characteristics and both have remarkable effects on ANNs. Gaussian noise is smooth and continuous, which works as a regularizer for artificial neural networks. On the other hand, Chaotic noise is irregular and also unpredictable and that works as a stimulus. Both the application of noises was compared to the baseline ANN on real data sets. The assessment of the accuracy and robustness of ANN performance under various types and amounts of noise was done. It was demonstrated that the noise profiling approach outperforms the baseline approach. It also examined the impact of Gaussian and Chaotic noise on the internal dynamics and representations of ANNs, providing some intriguing new information on how noise can affect ANN functionality and behaviour. In this research, two datasets were used: Animal and Shaded. The results demonstrated that bio-inspired noise profiling techniques can offer a straightforward yet efficient means of improving ANN performance for insect perception issues as well as diminish the overfitting of the model.