CS 224N Final Project

Motivated by the future need of intelligent systems for vehicle-to-passenger communication in autonomous vehicles, we implement two models for automatic image caption generation based on the neural encoder-decoder framework.

Following an extensive hyperparameter search, our non-attention based model achieves performance on the MSCOCO dataset highly comparable to that of its reference model. Our attention based model does however fail to exceed this performance, despite the attention mechanism appearing to function as intended. This leads to a discussion of possible causes and what architecture modifications that might be needed.