One-Shot Learning for Archaeological Site Data Using Deep Neural Network on Embedded Systems

Training Machine Learning models with effective features can be challenging and computationally expensive, especially when dealing with limited data. Conventional gradient-based approaches for target recognition require a large amount of data and extensive training. However, this study investigates a method for training Siamese Neural Networks using One-Shot learning, which reduces training time. By leveraging a unique structure, these networks can rank similarities between inputs and generalize to entirely new classes from unknown distributions. The Indian Archaeological Dataset is used to train and test the model, and after fine-tuning the network, its powerful discriminative features can be utilized to extend the predictive capabilities of the model beyond new data. In contrast, the Siamese Neural Network approach allows for effective feature extraction and recognition with limited data and reduces the risk of misconfiguring feature structures. Using this approach, we are able to achieve better accuracy of 86% and 66%, respectively, for One-Shot classification tasks for MNIST and Indian archaeological dataset by usage of triplet loss function. Using quantization, the model size is reduced to make it more compatible for working on embedded devices for classification tasks.