With machine learning (ML) technologies rapidly expanding to new applications and domains, users are collaborating with artificial intelligence-assisted diagnostic tools to a larger and larger extent. But what impact does ML aid have on cognitive performance, especially when the ML output is not always accurate? Here, we examined the cognitive effects of the presence of simulated ML assistance—including both accurate and inaccurate output—on two tasks (a domain-specific nuclear safeguards task and domain-general visual search task). Patterns of performance varied across the two tasks for both the presence of ML aid as well as the category of ML feedback (e.g., false alarm). These results indicate that differences such as domain could influence users’ performance with ML aid, and suggest the need to test the effects of ML output (and associated errors) in the specific context of use, especially when the stimuli of interest are vague or ill-defined
Eye tracking is a useful tool for studying human cognition, both in the laboratory and in real-world applications. However, there are cases in which eye tracking is not possible, such as in high-security environments where recording devices cannot be introduced. After facing this challenge in our own work, we sought to test the effectiveness of using artificial foveation as an alternative to eye tracking for studying visual search performance. Two groups of participants completed the same list comparison task, which was a computer-based task designed to mimic an inventory verification process that is commonly performed by international nuclear safeguards inspectors. We manipulated the way in which the items on the inventory list were ordered and color coded. For the eye tracking group, an eye tracker was used to assess the order in which participants viewed the items and the number of fixations per trial in each list condition. For the artificial foveation group, the items were covered with a blurry mask except when participants moused over them. We tracked the order in which participants viewed the items by moving their mouse and the number of items viewed per trial in each list condition. We observed the same overall pattern of performance for the various list display conditions, regardless of the method. However, participants were much slower to complete the task when using artificial foveation and had more variability in their accuracy. Our results indicate that the artificial foveation method can reveal the same pattern of differences across conditions as eye tracking, but it can also impact participants’ task performance.
Due to their recent increases in performance, machine learning and deep learning models are being increasingly adopted across many domains for visual processing tasks. One such domain is international nuclear safeguards, which seeks to verify the peaceful use of commercial nuclear energy across the globe. Despite recent impressive performance results from machine learning and deep learning algorithms, there is always at least some small level of error. Given the significant consequences of international nuclear safeguards conclusions, we sought to characterize how incorrect responses from a machine or deep learning-assisted visual search task would cognitively impact users. We found that not only do some types of model errors have larger negative impacts on human performance than other errors, the scale of those impacts change depending on the accuracy of the model with which they are presented and they persist in scenarios of evenly distributed errors and single-error presentations. Further, we found that experiments conducted using a common visual search dataset from the psychology community has similar implications to a safeguards- relevant dataset of images containing hyperboloid cooling towers when the cooling tower images are presented to expert participants. While novice performance was considerably different (and worse) on the cooling tower task, we saw increased novice reliance on the most challenging cooling tower images compared to experts. These findings are relevant not just to the cognitive science community, but also for developers of machine and deep learning that will be implemented in multiple domains. For safeguards, this research provides key insights into how machine and deep learning projects should be implemented considering their special requirements that information not be missed.
Deep Learning computer vision models require many thousands of properly labelled images for training, which is especially challenging for safeguards and nonproliferation, given that safeguards-relevant images are typically rare due to the sensitivity and limited availability of the technologies. Creating relevant images through real-world staging is costly and limiting in scope. Expert-labeling is expensive, time consuming, and error prone. We aim to develop a data set of both realworld and synthetic images that are relevant to the nuclear safeguards domain that can be used to support multiple data science research questions. In the process of developing this data, we aim to develop a novel workflow to validate synthetic images using machine learning explainability methods, testing among multiple computer vision algorithms, and iterative synthetic data rendering. We will deliver one million images – both real-world and synthetically rendered – of two types uranium storage and transportation containers with labelled ground truth and associated adversarial examples.
As the ability to collect and store data grows, so does the need to efficiently analyze that data. As human-machine teams that use machine learning (ML) algorithms as a way to inform human decision-making grow in popularity it becomes increasingly critical to understand the optimal methods of implementing algorithm assisted search. In order to better understand how algorithm confidence values associated with object identification can influence participant accuracy and response times during a visual search task, we compared models that provided appropriate confidence, random confidence, and no confidence, as well as a model biased toward over confidence and a model biased toward under confidence. Results indicate that randomized confidence is likely harmful to performance while non-random confidence values are likely better than no confidence value for maintaining accuracy over time. Providing participants with appropriate confidence values did not seem to benefit performance any more than providing participants with under or over confident models.