Sign in Subscribe
Artificial neural networks

Samsung Electronics Granted Patent for Low-Power Neural Network Inference Technique

Strategic Insights

— 1 min read

Samsung Electronics has been assigned a patent for a low-power change-based neural network inference technique for image processing, developed by four inventors: Ishay Goldin, Yonatan Dinai, Ran Vitek, and Michael Dinerstein. The patented technique enables high accuracy computer vision and image processing with decreased system resource requirements. One or more aspects of the present disclosure leverage key layers and compressed tensor comparisons to efficiently exploit temporal redundancy in videos and other slow-changing signals, reducing neural network inference computational burden.

Key Takeaways:

  • The patent (US 12354341 B2) was initially filed on May 20, 2022, by four inventors from Tel-Aviv, Israel: Ishay Goldin, Yonatan Dinai, Ran Vitek, and Michael Dinerstein.
  • The technique enables high accuracy computer vision and image processing with decreased system resource requirements, such as decreased computational load and shallower neural network designs.
  • Key aspects of the technique involve leveraging key layers and compressed tensor comparisons to efficiently exploit temporal redundancy in videos and other slow-changing signals.
  • The described techniques enable efficient reduction of neural network inference computational burden with only a minor increase in data transfer power consumption.
  • The technique can identify key layers of a neural network and efficiently leverage temporal/spatial redundancy across frames to reduce computation requirements.
  • Remaining feature-map calculations may be disabled in the layers between the identified key layers to further reduce computational burden.

Statistics:

  • The patent was initially filed on May 20, 2022 (Source: Samsung Electronics).
  • The technique enables a reduction in neural network inference computational burden with a minor increase in data transfer power consumption (Source: Samsung Electronics).
  • The technique can efficiently exploit temporal redundancy in videos and other slow-changing signals to reduce computational requirements (Source: Samsung Electronics).
  • The described techniques may leverage key layers and compressed tensor comparisons to reduce neural network inference computational burden (Source: Samsung Electronics).

Sources:

  • Samsung Electronics Co., Ltd., patent application (US 12354341 B2), initially filed on May 20, 2022.

Read more