This AI tool explains how AI “sees” images and why it can mistake an astronaut for a shovel

https://openaccess.thecvf.com/content/CVPR2023/papers/Fel_CRAFT_Concept_Recursive_Activation_FacTorization_for_Explainability_CVPR_2023_paper.pdf

It is widely acknowledged that artificial intelligence (AI) has made significant strides in recent years, resulting in significant breakthroughs and breakthroughs. However, it is not true that AI can achieve equally impressive results in all tasks. For example, while AI can outperform humans in certain visual tasks, such as face recognition, it can also make confusing errors in image processing and classification, highlighting the complex nature of the task. As a result, understanding the inner workings of such systems for a given task and how they make certain decisions has become a subject of great interest and research among researchers and developers. Similar to the human brain, AI systems are known to use image analysis and categorization strategies. However, the exact mechanisms of these processes remain elusive, resulting in a black box model.

Thus, there is a growing demand for explainability methods to interpret the decisions made by modern machine learning models, especially neural networks. In this context, attribution methods that generate heat maps that indicate the importance of individual pixels in influencing the model decision have gained popularity. However, recent research has revealed the limitations of these methods, as they tend to focus only on the most visible regions of the image, revealing where the model looks unclear what the model captures these areas. Thus, to demystify deep neural networks and reveal the strategies AI systems use to process images, a team of researchers from Brown University’s Carney Brain Science Institute and some computer scientists from the Toulouse Institute of Artificial and Natural Intelligence in France collaborated to develop. CRAFT (Concept Recursive Activation Factorization for Explainability). This innovative tool aims to distinguish “what” and “where” an AI model focuses on in the decision-making process, thus highlighting the differences in how the human brain and the computer vision system perceive visual information. The research was also presented at the prestigious Computer Vision and Pattern Recognition Conference 2023 held in Canada.

As mentioned earlier, it has been difficult to understand how AI systems make decisions using specific image regions using attribution techniques. However, simply identifying the influential regions without clarifying why these regions are crucial does not provide a comprehensive explanation for humans. CRAFT overcomes this limitation by using advanced machine learning techniques to unravel the complex and multidimensional visual representations learned by neural networks. To improve understanding, researchers have developed a user-friendly website where individuals can effortlessly explore and visualize these basic concepts that neural networks use to classify objects. In addition, the researchers also emphasized that when implementing CRAFT, users not only gain insight into the concepts that the AI ​​system uses to create an image and understand what the model perceives in specific domains, but also understand the hierarchical classification of these concepts. This revolutionary breakthrough offers a valuable resource for unraveling the decision-making process of AI systems and improving the transparency of their classification results.

🚀 Check out 100s of AI tools in our AI Tools Club

Basically, the main contribution of the work done by the researchers can be summarized in three main points. First, the team developed a recursive approach to efficiently identify and decompose concepts across multiple layers. This innovative strategy provides a comprehensive understanding of the underlying components of a neural network. Second, a revolutionary method is introduced to accurately assess the importance of concepts using Sobol indices. Finally, the introduction of implicit differentiation has revolutionized the creation of concept attribution maps, opening up a powerful tool to visualize and understand the relationship between concepts and pixel-level features. In addition, the team conducted a series of experimental evaluations to substantiate the effectiveness and relevance of their approach. The results revealed that CRAFT outperforms all other attribution methods, cementing its considerable utility and becoming a springboard for further research into concept-based explainability methods.

The researchers also stressed the importance of understanding how computers perceive images. By gaining deep insight into the visual strategies used by AI systems, researchers gain a competitive advantage by improving the accuracy and performance of vision-based tools. Additionally, this understanding is proving useful against adversaries and cyberattacks, helping researchers understand how attackers can fool AI systems by subtly changing pixel intensities in ways that are barely perceptible to humans. As for future work, the researchers are excited for the day when computer vision systems can surpass human capabilities. With the potential to address unsolved problems such as cancer diagnosis, fossil recognition, etc., the team strongly believes that these systems hold the promise of revolutionizing many fields.


Check Paper and reference article. Don’t forget to join our 25,000+ ML SubReddit, Discord channeland Email newsletter, where we share the latest AI research news, great AI projects, and more. If you have any questions about the above article or if we missed something, please email us at Asif@marktechpost.com


Featured tools:

Khushboo Gupta is an intern at MarktechPost Consultants. She is currently pursuing her Bachelor’s degree at the Indian Institute of Technology (IIT), Goa. She is passionate about machine learning, natural language processing and web development. She likes to learn more about the technical field by participating in several challenges.

🔥 StoryBird.ai just dropped some amazing features. Create an illustrated story from the prompt. Check it out here. (Sponsored)

#tool #explains #sees #images #mistake #astronaut #shovel

Leave a Comment