The problems we tackle

In order to bring deep learning visual perception AI to edge devices in a commercially viable manner, we have to enable our neural network based software to run

  • ⇒ in limited processor and memory environments
  • ⇒ with limited energy supply
  • ⇒ in real-time
  • ⇒ without losing significantly in accuracy

 

In order to bring deep learning visual perception AI to edge devices in a commercially viable manner, we have to enable our neural network based software to run

  • ⇒ in limited processor and memory environments
  • ⇒ with limited energy supply
  • ⇒ in real-time
  • ⇒ without losing significantly in accuracy

Our solution

Deep-learning visual perception solutions
that are well integrated with the hardware components
,
tailored to their use-cases and that make use of the latest advances
in network pruning and quantization techniques.

Deep-learning visual perception solutions
that are well integrated with the hardware components
,
tailored to their use-cases and that make use of the latest advances
in network pruning and quantization techniques.

We develop and innovate a variety of techniques that enable us to greatly reduce the memory, runtime and energy requirements of deep learning based image recognition (by more than 90%) without significantly sacrificing accuracy (~1 to 3%).

At the same time, our approach allows us to minimize the effort needed for annotating training data for our deep learning algorithms, which in turn makes new commercial applications profitable and thus commercially viable.