Michael Cogswell’s research while affiliated with SRI International and other places

Vision-language models (VLMs) have recently demonstrated strong efficacy as visual assistants that can parse natural queries about the visual content and generate human-like outputs. In this work, we explore the ability of these models to demonstrate human-like reasoning based on the perceived information. To address a crucial concern regarding the extent to which their reasoning capabilities are fully consistent and grounded, we also measure the reasoning consistency of these models. We achieve this by proposing a chain-of-thought (CoT) based consistency measure. However, such an evaluation requires a benchmark that encompasses both high-level inference and detailed reasoning chains, which is costly. We tackle this challenge by proposing a LLM-Human-in-the-Loop pipeline, which notably reduces cost while simultaneously ensuring the generation of a high-quality dataset. Based on this pipeline and the existing coarse-grained annotated dataset, we build the CURE benchmark to measure both the zero-shot reasoning performance and consistency of VLMs. We evaluate existing state-of-the-art VLMs, and find that even the best-performing model is unable to demonstrate strong visual reasoning capabilities and consistency, indicating that substantial efforts are required to enable VLMs to perform visual reasoning as systematically and consistently as humans. As an early step, we propose a two-stage training framework aimed at improving both the reasoning performance and consistency of VLMs. The first stage involves employing supervised fine-tuning of VLMs using step-by-step reasoning samples automatically generated by LLMs. In the second stage, we further augment the training process by incorporating feedback provided by LLMs to produce reasoning chains that are highly consistent and grounded. We empirically highlight the effectiveness of our framework in both reasoning performance and consistency.

We present a novel framework for probing and improving relational, compositional and contextual understanding of large visual-language models (V+L). While large V+L models have achieved success in various downstream tasks, it is not clear if they have a conceptual grasp of the content. We propose a novel benchmarking dataset for probing three aspects of content understanding. Our probes are grounded in cognitive science and help determine if a V+L model can, for example, determine if snow garnished with a man is implausible, or if it can identify beach furniture by knowing it is located on a beach. We have experimented with 5 well known models, such as CLIP and ViLT, and found that they mostly fail to demonstrate a conceptual understanding. That said, we find interesting insights such as cross-attention helps learning conceptual understanding. We use these insights to propose a new finetuning technique that rewards the three conceptual understanding measures we proposed. We hope that the presented benchmarks will help the community assess and improve the conceptual understanding capabilities of large V+L models.

Attention maps, a popular heatmap‐based explanation method for Visual Question Answering (VQA), are supposed to help users understand the model by highlighting portions of the image/question used by the model to infer answers. However, we see that users are often misled by current attention map visualizations that point to relevant regions despite the model producing an incorrect answer. Hence, we propose Error Maps that clarify the error by highlighting image regions where the model is prone to err. Error maps can indicate when a correctly attended region may be processed incorrectly leading to an incorrect answer, and hence, improve users’ understanding of those cases. To evaluate our new explanations, we further introduce a metric that simulates users’ interpretation of explanations to evaluate their potential helpfulness to understand model correctness. We finally conduct user studies to see that our new explanations help users understand model correctness better than baselines by an expected 30% and that our proxy helpfulness metrics correlate strongly (𝜌 > 0.97) with how well users can predict model correctness.

In the domain of Visual Question Answering (VQA), studies have shown improvement in users’ mental model of the VQA system when they are exposed to examples of how these systems answer certain Image-Question (IQ) pairs¹. In this work, we show that showing controlled counterfactual image-question examples are more effective at improving the mental model of users as compared to simply showing random examples. We compare a generative approach and a retrieval-based approach to show counterfactual examples. We use recent advances in generative adversarial networks (GANs) to generate counterfactual images by deleting and inpainting certain regions of interest in the image. We then expose users to changes in the VQA system’s answer on those altered images. To select the region of interest for inpainting, we experiment with using both human-annotated attention maps and a fully automatic method that uses the VQA system’s attention values. Finally, we test the user’s mental model by asking them to predict the model’s performance on a test counterfactual image. We note an overall improvement in users’ accuracy to predict answer change when shown counterfactual explanations. While realistic retrieved counterfactuals obviously are the most effective at improving the mental model, we show that a generative approach can also be equally effective.

Despite their success and popularity, deep neural networks (DNNs) are vulnerable when facing backdoor attacks. This impedes their wider adoption, especially in mission critical applications. This paper tackles the problem of Trojan detection, namely, identifying Trojaned models -- models trained with poisoned data. One popular approach is reverse engineering, i.e., recovering the triggers on a clean image by manipulating the model's prediction. One major challenge of reverse engineering approach is the enormous search space of triggers. To this end, we propose innovative priors such as diversity and topological simplicity to not only increase the chances of finding the appropriate triggers but also improve the quality of the found triggers. Moreover, by encouraging a diverse set of trigger candidates, our method can perform effectively in cases with unknown target labels. We demonstrate that these priors can significantly improve the quality of the recovered triggers, resulting in substantially improved Trojan detection accuracy as validated on both synthetic and publicly available TrojAI benchmarks.