Deep Learning Optimizer

What is the Deep Learning Optimizer?

For many applications, especially in the embedded systems area, this has long been a hurdle. Today, through the targeted use of arithmetic and topological optimization methods (quantization and pruning), the complexity can (usually) be reduced many times over without negative effects on the quality.

What is the EYYES Deep Learning Optimizer doing?

The Deep Learning Optimizer receives an already trained network. The complexity of the model is reduced by means of so-called post-training optimization procedures. An automated evaluation shows the achieved reduction. If use-case image data is available, the evaluation also shows the quality results compared to the unoptimized model. As a result, the optimized model and a binary model for the EYYES Deep Learning Accelerator are returned.

How does the EYYES Deep Learning Optimizer work?

The Deep Learning Optimizer receives an already trained network. The complexity of the model is reduced by means of so-called post-training optimization procedures. An automated evaluation shows the achieved reduction. If use-case image data is available, the evaluation also shows the quality results compared to the unoptimized model. The result is an optimized model and a binary model as input for the EYYES Deep Learning Accelerator.

Benefit and Advantage of the Optimizer

Neural network optimization allows cheaper or more energy-efficient hardware to be used for the application, such as our EYYES Deep Learning Accelerator. Post-training optimization allows the costly and time-consuming model training to take place normally:

  • No costly experiments during the training process
  • Full focus on quality
  • No additional image data is needed for optimization
  • 17-50% lower quantization losses