Understanding Visual Concepts Across Models

Brandon Trabucco, Max A Gurinas, Kyle Doherty, Russ Salakhutdinov

Keywords: Deep Learning, Prompt Tuning

Abstract: Large multimodal models such as Stable Diffusion can generate, detect, and classify new visual concepts after fine-tuning just a single word embedding. Do models learn similar words for the same concepts (i.e. \texttt{<orange-cat>} = orange + cat)? We conduct a large-scale analysis on three state-of-the-art models in text-to-image generation, open-set object detection, and zero-shot classification, and find that new word embeddings are model-specific and non-transferable. Across 4,800 new embeddings trained for 40 diverse visual concepts on four standard datasets, we find perturbations within an $\epsilon$-ball to any prior embedding that generate, detect, and classify an arbitrary concept. When these new embeddings are spliced into new models, fine-tuning that targets the original model is lost. We show popular soft prompt-tuning approaches find these perturbative solutions when applied to visual concept learning tasks, and embeddings for visual concepts are not transferable. Code for reproducing our work is available at: \href{https://anonymous-visual-words.github.io}{anonymous-visual-words.github.io}.

Submission Number: 7