Towards Sparse, Efficient and Explainable Data Attribution in Large AI Models

PDF view of the paper entitled Dualxda: Towards the support of scattered, effective and interpretable data in the large AI models, by Galip \ “UMIT Yolcu and 4 other authors

PDF HTML (experimental) view

a summary:Deep learning models achieve significant performance, however decision -making processes often remain transparent. In response, the field of interpretative artificial intelligence (Xai) has grown significantly over the past decade, mainly focusing on ways to support features. To complete this perspective, data support (DA) has emerged as a promising model that converts the focus from features to the data source. However, the current DA methods suffer from the high arithmetic costs and memory requirements. In addition, current support methods show a low decrease, which impedes the discovery of decisive patterns in data. We offer Dualxda, a scattered, effective and interpretative DA frame, consisting of interlocking approaches to support dual data (Dualda) and interpretable data assignment (XDA): with Dualda, we suggest effective and effective DA, taking advantage of the theory of the supportive support machine naturally and naturally contesting. We explain that Dualda achieves a high quality of support, and outperforms the solution of a series of assessed tasks, while improving the time of interpretation at the same time with a factor of up to 4,100,000 dollars \ $ compared to the original impact function method, and up to 11 thousand $ / $ compared to the most effective conjugation of literature. We also offer XDA, which is a way to enhance data assignment with capabilities of advantages of features to explain the reason for training samples documents to predict the test sample in terms of influencing features. Combated, our contributions to Dualxda ultimately indicate a future of artificial intelligence that is unprecedented, which allows a transparent, effective and new analysis even for the largest nerve structure that enhances a new generation of artificial intelligence. This URL https symbol.