Hybrid Embedding and Transformer-based Text and Vision Architectures for Deep Learning Artificial Intelligence

Deep learning architectures are able to efficiently extract very deep contextual information from large amounts of complex data in either vision or natural language. This study examines the performance of transformer-based models versus traditional deep learning models (CNN, RNN, GRU, LSTM, BiLSTM), comparing them using hybrid representation methods for both text and image datasets. Experiments were conducted on well-known text datasets (AG News and Yelp Polarity), as well as a subset of an image dataset (CIFAR-10 and ImageNet). Overall, experiments demonstrated that the performance of all transformer-based models consistently surpassed that of traditional deep learning models; moreover, when ensembling these two different types of models, their combined performance was even better. The performance/computational complexity/interpretability trade-off. In addition, this research demonstrates that the use of hybrid embeddings can significantly enhance the ability of models to function adequately in under-resourced scenarios. Furthermore, ensembling models that incorporate both classical deep learning models and transformer-based models will produce superior and generalized performance across various tasks.