Chameleon: Adaptive Code Optimization for Expedited Deep Neural Network Compilation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Current neural network execution methods rely on suboptimal, time-consuming techniques like hand-optimized libraries or genetic algorithms.
• These methods often involve frequent, costly hardware measurements, hindering efficiency and innovation.

Purpose of the Study:

• To develop an adaptive solution for rapid code optimization in neural networks.
• To accelerate the search for optimal code and enhance output performance for unseen design spaces.

Main Methods:

• Leveraging reinforcement learning (RL) for faster convergence in optimization.
• Developing an adaptive sampling algorithm that prioritizes representative hardware measurements.
• Incorporating domain-knowledge inspired logic to improve sample quality.

Main Results:

• Chameleon achieved a 4.45x speed-up in optimization time compared to AutoTVM.
• Demonstrated a 5.6% improvement in inference time for modern deep networks.
• Successfully adapted to previously unseen design spaces for efficient code optimization.

Conclusions:

• Chameleon offers a more efficient and effective approach to neural network code optimization.
• The adaptive sampling and RL integration significantly reduce optimization time and improve performance.
• This method fosters innovation by enabling faster iteration and deployment of diverse neural network architectures.