One of the observations that intrigued me the most during the experiment with Leonardo was the difficulties that AI encountered when it is given negative instructions. While attempting to construct a self portrait using Leonardo, I used descriptions such as “a girl not wearing glasses”. To my surprise, almost all the responses I received were a girl wearing glasses. I then tried another prompt using similar negative instructions, such as “a girl not smiling”; and again, AI’s responses were also opposite to the instructions.
It immediately made me think of “The White Bear Problem”, which is describing the phenomenon that when a person is given the instruction to not think about a white bear, the first thing that comes to their head is the white bear. I then thought, do AI models have similar problems as well? Or maybe, rather than thinking of the white bear first and then executing the negative response like a human, AI is only capable of doing the first step i.e. responding with an image of a white bear?
I then conducted some research regarding these problems. According to a study about large multi-modal models (LMMS) and Hallucinations, AI models are more commonly trained to understand positive instructions, and thus they are inherently less good at processing negative instructions (Liu et al., 2024). They might also “over-rely on language priors and generate words more likely to go together with the instruction text regardless of the image content” (Liu et al., 2024, p.1), which suggests that AI models do not generate content through accurately interpreting the meaning of the instructions like humans do, but through analysing the language patterns and generate with a response that has the most repetitions with the description.
This paper also discussed another type of negative instruction called “Existent Object Manipulation” (Liu et al., 2024), which is about providing inconsistent attributes of an existing object in the image. For example, asking the model about a “woman in blue pants” when the woman is in fact wearing red pants in the image. I think this might be relevant to another frustration I encountered while using Leonardo: After providing me an image of students in darker skin colours, I asked it to produce another image with students from different races. Somehow, it failed to achieve that. Is this an example of the incomplete training of existent object manipulation? I think this is a question that can only be answered after I read more papers.
Overall, although the experiment with Leonardo has given me frustrations, it has certainly triggered some interest in me to study more about the mechanisms of AI and how scientists can train different models.
References:
Liu, F., Lin, K., Li, L., et al., 2024, ‘Mitigating hallucination in large multi-modal models via robust instruction tuning’, International Conference on Learning Representations (ICLR), Ithaca, accessed 31 August 2024, ProQuest One Academic database.