Neural Architecture Selection as a Nash Equilibrium With Batch Entanglement.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Differentiable Neural Architecture Search (DARTS) methods face challenges in discretizing one-shot architectures.
• Existing heuristic or progressive search methods are inefficient and prone to local optima.

Purpose of the Study:

• To develop a novel and efficient method for discretizing and selecting single-path architectures from one-shot models.
• To address the limitations of current DARTS discretization techniques.

Main Methods:

• Formulating architecture search as an 'architecture game' with 'keep' and 'drop' strategies.
• Identifying the Nash equilibrium of the game to extract the optimal single-path architecture.
• Employing entangled Gaussian representation of mini-batches inspired by Parrondo's paradox for enhanced efficiency.

Main Results:

• The proposed method significantly accelerates the discretization process compared to state-of-the-art techniques.
• The approach achieves competitive performance with higher maximum accuracy.
• Demonstrated effectiveness through extensive experiments on benchmark datasets.

Conclusions:

• The architecture game framework provides an effective and efficient solution for DARTS discretization.
• The novel method overcomes the limitations of previous approaches, offering improved speed and accuracy.
• This work advances the field of automated machine learning by refining architecture search strategies.