[Paper Review] Examining Cooperation in Visual Dialog Models

Mircea Mironenco, Dana Kianfar+3 more|arXiv (Cornell University)|Dec 4, 2017

Multimodal Machine Learning Applications10 references3 citations

TL;DR

This paper introduces a black-box intervention method to assess the contribution of linguistic and visual components in visual dialog models. By systematically impairing image features, captions, or dialog turns—via random noise or negation—it reveals that state-of-the-art models rely minimally on visual input and dialog coherence, instead depending heavily on initial captions, indicating weak cooperation between agents.

ABSTRACT

In this work we propose a blackbox intervention method for visual dialog models, with the aim of assessing the contribution of individual linguistic or visual components. Concretely, we conduct structured or randomized interventions that aim to impair an individual component of the model, and observe changes in task performance. We reproduce a state-of-the-art visual dialog model and demonstrate that our methodology yields surprising insights, namely that both dialog and image information have minimal contributions to task performance. The intervention method presented here can be applied as a sanity check for the strength and robustness of each component in visual dialog systems.

Motivation & Objective

• To develop a model-agnostic method for evaluating the contribution of individual components—linguistic and visual—in visual dialog systems.
• To investigate whether visual dialog agents truly cooperate by leveraging both image and language information during conversation.
• To identify whether models over-rely on initial captions or exploit dialog interactions effectively for image retrieval.
• To provide a sanity check for model robustness and component-level reliability in multimodal dialog systems.

Proposed method

• Apply structured or randomized interventions during inference to impair specific components: image features, captions, or dialog turns.
• Replace image features with random uniform noise to test visual modality dependence.
• Replace content words in captions with random words to assess caption quality and relevance.
• Randomly corrupt tokens in questions or answers with probability p to evaluate dialog robustness.
• Use negation intervention by flipping yes/no answers to test sensitivity to non-cooperative behavior.
• Evaluate performance using mean-percentile rank (MPR) on the VisDial validation set under each intervention condition.

Experimental results

Research questions

• RQ1 To what extent does the visual dialog model depend on visual input for accurate image retrieval?
• RQ2 How much does the model rely on the initial caption compared to dialog history for making predictions?
• RQ3 How robust is the model to random or adversarial perturbations in dialog turns?
• RQ4 Does the model exhibit cooperative behavior when faced with negated responses from the answer bot?
• RQ5 Are linguistic and visual components equally exploited, or is one dominant in the decision-making process?

Key findings

• Replacing image features with random noise caused only a 0.4% drop in MPR, indicating minimal reliance on visual input.
• Caption interventions led to a 40.1% performance drop at round 10, showing that initial captions are the primary source of information.
• Randomly corrupting answers or questions caused less than a 1% MPR drop, suggesting dialog content is not effectively used for image identification.
• Negation interventions caused only a 0.3% performance drop, indicating limited sensitivity to non-cooperative behavior.
• The model's performance remained stable under most interventions, revealing weak inter-agent cooperation and potential overfitting to initial captions.
• The model’s ranking performance improved slightly over time under caption intervention, but never recovered, indicating persistent degradation from corrupted initial context.