What is it about?
Deep Neural Networks (DNNs) have gained attention in domain-specific supervised learning, but face challenges in acquiring well-labeled training data and training efficiency due to imprecise object characteristics. We propose a formal specification-based framework for DNNs, defining important features to generate training and testing data.
Featured Image
Photo by Christopher Gower on Unsplash
Why is it important?
We consider the assets of formal specifications and propose a framework of a method for automatically generating labeled data based on formal specifications. Since the foundation of data generation in our framework is the formal specification, apart from the presentation of the whole method, this work also focuses on the discussion of how a formal specification can be properly written to reflect the distinct features of the object to be recognized by DNNs. The method presented in this paper can provide a new perspective on the data source of DNNs, especially for object recognition.
Perspectives
In this work, we propose a method for generating training data for supervised learning based on SOFL formal specifications, trying to provide a solution to the difficulty of obtaining well-labeled training data. In particular, we concentrate on the discussion of the details about how to use SOFL specifications to define the characteristics of the objects to be classified by a DNN model. The detailed steps of this framework still need some further research. But we believe it would be a good solution to the data-obtaining issue for DNNs.
YANZHAO XIA
Hiroshima Daigaku
Read the Original
This page is a summary of: A Framework of Formal Specification-Based Data Generation for Deep Neural Networks, February 2023, ACM (Association for Computing Machinery),
DOI: 10.1145/3587828.3587869.
You can read the full text:
Resources
Contributors
The following have contributed to this page
