Overcoming Bias in Captioning Models

Most machine learning models are known to capture and exploit bias. While this can be beneficial for many classification tasks (e.g., it might be easier to recognize a computer mouse given the context of a computer and a desk), exploiting bias can also lead to incorrect predictions. In this talk, I will first consider how over-reliance on bias might lead to incorrect predictions in a scenario where is inappropriate to rely on bias: gender prediction in image captioning. I will present the Equalizer model which more accurately describes people and their gender by considering appropriate gender evidence. Next, I will consider how bias is related to hallucination, an interesting error mode in image captioning. I will present a metric designed to measure hallucination and consider questions like what causes hallucination, which models are prone to hallucination, and do current metrics accurately capture hallucination?