Artificial neural networks are typically initialized using mathematically defined methods that do not account for the structural and functional diversity of biological systems. While traditional approaches ensure training stability, they overlook natural mechanisms of synaptic connection formation. This study proposes a biologically inspired approach to weight initialization based on stochastic patterns identified in empirical movement trajectories recorded in a controlled biological environment. The data undergo preprocessing, including smoothing, normalization, and scaling, to generate weight values subsequently used for neural network initialization. The effectiveness of the proposed method is evaluated against standard initialization strategies using three benchmark datasets: MNIST, Fashion-MNIST, and the Gas Sensor Array Drift dataset. Experimental results demonstrate that the biologically inspired approach achieves comparable performance across all evaluation criteria, including test and validation accuracy, the number of epochs required for convergence, class-wise sensitivity, and the macro-averaged F1-score. In several cases, the method facilitates faster convergence without compromising classification accuracy. Although the proposed strategy does not consistently outperform conventional methods, it introduces structured stochasticity into the training process based on biological principles and provides a promising foundation for further research into more complex architectures and biologically inspired learning models.
Zolotukhin et al. (Mon,) studied this question.