This work deals with the automated classification of Celtic coins using neural networks. Celtic coins are historical artifacts with unique characteristics, and their classification provides archaeologists and historians with deeper insights into their origin and significance. Given the complexity of manual classification, neural networks offer a promising alternative for efficiently processing large datasets. At the core of the study was the comparison of different network architectures and input data. Four approaches were examined: (1) using the obverse sides of the coins as the sole data input, (2) using the reverse sides of the coins as the sole data input, (3) using both the obverse and reverse sides on a single image as the sole data input, and (4) using both the obverse and reverse sides of the coins as separate data inputs in parallel Convolutional Neural Networks (CNNs) connected via a Concat layer. The results showed that the model trained exclusively on the reverse sides of the coins achieved the best predictions. The multi-input model that utilized both sides of the coins as separate inputs emerged as a promising alternative. Furthermore, it was observed that the reverse sides of the coins generally provided higher information content for classification. This work highlights the importance of careful selection of network architectures and input data. The findings suggest that when classifying Celtic coins, the use of separate inputs for obverse and reverse sides may potentially be advantageous.