Micro-architecture design exploration template for AutoML case study on SqueezeSEMAuto.

Area of Science:

• Deep Learning
• Computer Vision
• Machine Learning

Background:

• Convolutional Neural Network (CNN) models are vital for image recognition but require extensive manual tuning for optimal architecture.
• Automated Machine Learning (AutoML) offers a solution to streamline this process.

Purpose of the Study:

• To develop and evaluate an AutoML framework for exploring micro-architecture blocks and multi-input options in CNNs.
• To enhance SqueezeNet architectures using SE blocks and residual combinations for superior performance.

Main Methods:

• Exploited an AutoML framework incorporating Random, Hyperband, and Bayesian search strategies.
• Applied the framework to SqueezeNet, integrating SE blocks and residual combinations.
• Evaluated performance on CIFAR-10 and Tsinghua Facial Expression datasets.

Main Results:

• Achieved higher accuracy than traditional architectures without manual effort.
• Demonstrated significant accuracy gains on CIFAR-10 (up to 78% with SE blocks vs. 50% for traditional SqueezeNet).
• Showcased substantial improvements in facial expression recognition (up to 71% vs. under 20% for traditional models).

Conclusions:

• The proposed AutoML approach effectively automates CNN architecture optimization, yielding superior accuracy and manageable model sizes.
• This method significantly reduces the need for time-consuming manual experimentation in deep learning model design.