NAS-Navigator: Visual Steering for Explainable One-Shot Deep Neural Network Synthesis.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep learning (DL) success relies on expert tuning of parameters and architectures.
• Neural Architecture Search (NAS) automates DNN architecture design, aiding non-experts.
• One-shot NAS techniques reduce search time by training a single template network.

Purpose of the Study:

• To address the lack of explainability and human-in-the-loop (HIL) control in One-Shot NAS.
• To introduce NAS-Navigator, a visual analytics (VA) system for enhanced One-Shot NAS.
• To improve performance and reduce search time compared to existing state-of-the-art (SOTA) techniques.

Main Methods:

• Developed NAS-Navigator, a visual analytics system integrating HIL design and explainability into One-Shot NAS.
• Employed parameter sharing in a large template network to encompass all candidate DNNs.
• Utilized component ranking based on evaluating randomly selected candidate architectures.

Main Results:

• NAS-Navigator provides users with control over NAS while retaining automated search benefits.
• The improved One-Shot NAS algorithm demonstrates performance comparable to SOTA techniques.
• Integrating VA with NAS-Navigator further enhances search time and overall performance.

Conclusions:

• NAS-Navigator effectively improves explainability and HIL control for One-Shot NAS.
• The system empowers non-expert users by leveraging domain knowledge to guide the search.
• NAS-Navigator offers competitive performance and efficiency gains in automated DNN architecture search.