ABSTRACT
In today’s rapidly evolving artificial intelligence (AI) landscape, the complexity of neural networks, especially in deep learning, presents significant challenges for intuitive human understanding. Compelling visualization methods have become crucial with AI systems integrating into everyday experiences and environments. Virtual Reality (VR), as an immersive and interactive technology, offers a novel approach to visualizing intricate AI processes. This work introduces transformative updates to the DeepVisionVR platform, a pioneering tool for 3D visualization of Convolutional Neural Networks (CNNs) in VR. A cornerstone of our enhancements is the Sensitivity Analysis module, which offers real-time interactivity, allowing users to adjust pixels within the VR space, shedding light on the intricacies of model outcomes. In addition, advanced model interpretation methodologies have been integrated, including Integrated Gradients, GradientShap, and Occlusion, enriching the depth of insight into model rationale. Our Adversarial Analysis, utilizing the Fast Gradient Sign Method (FGSM), unveils potential weak points in models, emphasizing their vulnerability to minor input alterations. The new features aim to provide a deeper understanding of how neural networks interpret and react to various inputs, thereby bridging the gap between complex machine learning models and human interpretability.
- Medet Inkarbekov, Rosemary Monahan, and Barak A. Pearlmutter. 2023. Visualization of AI Systems in Virtual Reality: A Comprehensive Review. International Journal of Advanced Computer Science and Applications 14, 8 (2023). https://doi.org/10.14569/IJACSA.2023.0140805Google ScholarCross Ref
- Narine Kokhlikyan, Vivek Miglani, Miguel Martin, Edward Wang, Bilal Alsallakh, Jonathan Reynolds, Alexander Melnikov, Natalia Kliushkina, Carlos Araya, Siqi Yan, 2020. Captum: A unified and generic model interpretability library for pytorch. arXiv preprint arXiv:2009.07896 (2020).Google Scholar
- Christoph Linse, Hammam Alshazly, and Thomas Martinetz. 2022. A walk in the black-box: 3D visualization of large neural networks in virtual reality. Neural Computing and Applications 34, 23 (2022), 21237–21252. https://doi.org/10.1007/s00521-022-07608-4Google ScholarDigital Library
Index Terms
- Immersive Neural Network Exploration: A VR Approach to Human-Centered AI Understanding
Recommendations
Future Standards for Immersive VR: Report on the IEEE Virtual Reality 2007 Workshop
The Future Standards for Immersive VR Workshop at the IEEE Virtual Reality 2007 Conference aimed to foster an understanding between the VR and Web3D communities. Through multiple tool and application examples, contributors looked at the challenges and ...
Immersive virtual studio for architectural exploration
3DUI '10: Proceedings of the 2010 IEEE Symposium on 3D User InterfacesArchitects use a variety of analog and digital tools and media to plan and design constructions. Immersive virtual reality (VR) technologies have shown great potential for architectural design, especially for exploration and review of design proposals. ...
Exploration & Anthropomorphism in Immersive Unit Visualizations
CHI EA '18: Extended Abstracts of the 2018 CHI Conference on Human Factors in Computing SystemsWe report on an initial examination of the potential of immersive unit visualizations in virtual reality, showing how these visualizations can help viewers examine data at multiple scales and support affective, personal experiences with data. We outline ...
Comments