Smart but slow
Neural networks need not only solve complex problems. Most of the time they also need to run very fast. In this project the customer already had a big neural network that worked well for their use case. But it required powerful hardware, and they needed our help to improve the inference speed on edge devices.
Optimized for speed
We helped the customer utilize the latest hardware accelerations on the target edge devices and quantization solutions; to optimize and greatly improve the inference speed.
The project resulted in a much faster inference speed with a smaller model in size; without compromising model accuracy. The customer could now utilize the model on the target edge devices and provide the same experience as when the model was run on a much more powerful device.