ExplainabilityTransparencyAccountabilitycs.LG
Explainable Deep Learning for Video Recognition Tasks: A Framework & Recommendations
Liam Hiley, Alun Preece, Yulia Hicks
September 7, 2019
arXiv.org
16 citations
Key Takeaways
-
Video AI models are significantly more complex than image models, containing at least twice as many parameters, yet they receive far less attention in explainability research—creating a gap between the sophistication of these systems and our ability to understand how they make decisions.
-
Current explanation methods for video AI are borrowed from image recognition rather than purpose-built for video, which is problematic because video involves spatio-temporal (space and time) dimensions that still images do not, meaning existing explanations may not adequately capture how these models process motion and temporal sequences.
-
The explainability research landscape lags behind video AI development, with the paper noting a "scarcity of research into explanations for these methods," suggesting that as video recognition systems become more prevalent in real-world applications, we risk deploying increasingly opaque technology without adequate tools to audit or understand their behavior.
-
Purpose-built explainability methods for video are necessary for responsible deployment, particularly as these complex models are applied to high-stakes domains where understanding model decisions (such as surveillance, autonomous vehicles, or medical video analysis) becomes critical for accountability and trust.
Abstract
The popularity of Deep Learning for real-world applications is ever-growing. With the introduction of high performance hardware, applications are no longer limited to image recognition. With the introduction of more complex problems comes more and more complex solutions, and the increasing need for explainable AI. Deep Neural Networks for Video tasks are amongst the most complex models, with at least twice the parameters of their Image counterparts. However, explanations for these models are often ill-adapted to the video domain. The current work in explainability for video models is still overshadowed by Image techniques, while Video Deep Learning itself is quickly gaining on methods for still images. This paper seeks to highlight the need for explainability methods designed with video deep learning models, and by association spatio-temporal input in mind, by first illustrating the cutting edge for video deep learning, and then noting the scarcity of research into explanations for these methods.