Comparative evaluation of noise performance for motion recognition between animals and AI

Abstract

This study explores motion recognition in humans, fruit flies, and the 3DResNet neural network, focusing on their resistance to salt-and-pepper noise. Humans and fruit flies represent biological models, while 3DResNet represents a currrent DNN(Deep Neural Network). Motion recognition is essential for survival. Humans have possibly the highest level of cognitive ability in the animal kingdom, and typically set the benchmark for cognitive tasks. Fruit flies, with their simpler neural system, can provide insight into fundamental sensorimotor processes. Meanwhile, 3DResNet, a powerful neural network, is specialized for visual recognition. Our analysis goes beyond accuracy metrics, including a comparison of temporal dynamics where applicable. Humans exhibit high accuracy, while fruit flies surprisingly show faster response times, suggesting efficient sensorimotor integration. 3DResNet, despite lower accuracy with noiseless data, adapts well and outperforms both flies and humans when trained with noisy datasets. This study challenges expectations, suggesting a divergence in adaptation strategies to environmental noise. The results contribute insights into motion perception mechanisms in biological and artificial systems, guiding future research in neuroscience and artificial intelligence. The findings highlight the importance of considering varied adaptation mechanisms in the development of robust recognition algorithms.