Sign in Subscribe
Artificial intelligence

Optimizing Autotvm with Parallel Genetic Algorithms Boosts AI Performance

Strategic Insights

— 1 min read

Researchers from Beihang University in Beijing have made significant advancements in the field of machine learning, discovering a new method to optimize auto-tuning for deep neural networks. The team, led by Yuchen Feng, employed a parallel genetic algorithm (PGA) in conjunction with the AutoTVM auto-tuning process to achieve improved performance. According to the study, the new method reduces the inference time of typical deep networks by about 8-14% and speeds up the time consumption of the auto-tuning process up to 1.2-1.52x on GPUs compared to the original GA process of AutoTVM.

Key Takeaways:

  • The researchers combined a genetic algorithm (GA) with AutoTVM, resulting in improved performance and reduced inference time.
  • The new method achieves double optimization on both tuning results and tuning time.
  • The parallel genetic algorithm (PGA) was used to widen the search scope and improve search efficiency.
  • The study was conducted with the support of the National Engineering Research Center of Oil & Gas Exploration Computer Software Open Fund.
  • Beihang University's Yuchen Feng was the lead researcher on the project.
  • Additional authors included Changhai Zhao, Jiamin Wen, Minqiang Shang, and Xiaohua Shi.
  • The new method has been tested and shown to improve the inference time of typical deep networks by 8-14%.

Statistics:

  • Reduction in inference time: 8-14%
  • Speedup in auto-tuning process: up to 1.2-1.52x on GPUs
  • Number of researchers involved: 5 (Feng, Zhao, Wen, Shang, Shi)
  • Funding source: National Engineering Research Center of Oil & Gas Exploration Computer Software Open Fund
  • Journal publication: International Journal of Pattern Recognition and Artificial Intelligence (39, 2025, World Scientific Publ Co Pte Ltd)

Sources:

  • Optimizing Autotvm By Parallel Genetic Algorithms. International Journal of Pattern Recognition and Artificial Intelligence, 2025;39.
  • World Scientific Publishing - www.worldscientific.com/;
  • International Journal of Pattern Recognition and Artificial Intelligence - www.worldscinet.com/ijprai/ijprai.shtml
  • Yuchen Feng, Beihang University, School of Software, Beijing 100083, People's Republic of China.

Read more