An In-Situ Visual Analytics Framework for Deep Neural Networks.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Deep neural networks (DNNs) demonstrate significant power across domains but training them is complex due to massive parameters.
• Current visualization methods for DNNs rely on post-hoc analysis of logged data, which is inefficient for large datasets and complex models.
• Existing methods face challenges with large data storage, I/O overhead, and lack of real-time human intervention capabilities.

Purpose of the Study:

• To propose an in-situ visualization and analysis framework for DNN training.
• To address the limitations of traditional post-hoc analysis in DNN training.
• To enable real-time monitoring and intervention during DNN model development.

Main Methods:

• Implemented an in-situ visualization framework for DNN training.
• Utilized feature extraction algorithms for in-situ data reduction.
• Enabled real-time visual analytics and human intervention during the training process.

Main Results:

• The framework effectively reduces the size of training-related data in-situ.
• Real-time visual analytics provide immediate insights into model training states.
• Case studies demonstrate improved DNN optimization and analysis efficiency for deep learning experts.

Conclusions:

• The proposed in-situ framework enhances DNN training by enabling real-time analysis and intervention.
• This approach overcomes the scalability and efficiency issues of post-hoc visualization methods.
• The framework empowers deep learning experts to optimize models more effectively and efficiently.