Differentiating Human and Machine Intelligence with Contextualized Embeddings

Abstract

This thesis explores the application of deep learning models as contextual embedding functions for data enrichment in the Turing test, allowing an AI-based judge that automates the Turing test to more effectively differentiate between human and artificial intelligence inputs in tasks such as image captioning. Image captioning data were manually collected from both humans and AI models such as BLIP and GIT. Static embedding functions are first applied to the data before being propagated through pre-trained deep learning models of various architectures to obtain contextualized embeddings for classification. Specifically, transformers and convolutional neural networks are used to generate contextual embeddings for image and text data respectively. PCA dimensionality reduction is applied on the contextual embedding space to alleviate memory and resource constraints for training the AI-based judge for human-machine classification. AI judges based on support vector machines (SVM), Gaussian Naive Bayes and deep neural networks are trained and evaluated on the contextual embeddings, and the resulting performance metrics for classification accuracy are discussed. Further insights about possible innate differences between humans and AI in the domains of vision and language are analyzed.