Comparative Study of Activation Functions and Image Resolution on ResNet-34 for Spiral Galaxy Spin Classification

This study investigates the application of the Residual Network (ResNet-34) architecture for classifying spiral galaxy spin directions, specifically focusing on the comparative performance of activation functions and cross-dataset generalizability using data derived from the Dark Energy Spectroscopic Instrument Legacy Survey (DESI LS) and the Hyper Suprime Cam Subaru Strategic Program (HSC-SSP) surveys. The methodology ensures robustness by training each model configuration across 10 independent runs. The results demonstrate the clear superiority of the Rectified Linear Unit (ReLU) over the Hyperbolic Tangent (Tanh); ReLU-based models achieved a mean peak accuracy of 94.7% and required only less than 60 epochs to converge, significantly faster than Tanh's 120 epochs. Crucially, we found that models trained on lower-resolution DESI LS images exhibited superior robustness and generalizability compared to high-resolution-trained models, suggesting that low-resolution training acts as effective implicit regularization. This research provides critical design recommendations for efficient machine learning pipelines, particularly for upcoming facilities like the 3.8-meter telescope at Timau National Observatory (TNO), ensuring model stability and transferability across diverse survey conditions.