Sign in Subscribe
Technical institutes

Breakthrough in Sensor Research: Low-Cost and Highly Energy-Efficient Convolutional Neural Networks

Strategic Insights

— 1 min read

Researchers at Beijing Information Science and Technology University have made a significant advancement in sensor research, developing a low-cost and highly energy-efficient convolutional neural network (CNN) that can execute tasks at the near-sensor end. This innovative design, based on a hybrid encoding of deterministic and binary encoding, has the potential to revolutionize the field of machine learning and edge devices. By reducing the length of the bit stream input and optimizing the network training process, the researchers achieved a recognition rate of 99% with a 2-bit input, significantly reducing system latency and energy consumption.

Key Takeaways:

  • The research proposes a hardware optimization design of CNN based on the hybrid encoding of deterministic encoding and binary encoding, which achieves a low-cost and high-energy-efficiency convolution operation network.
  • The network can achieve good recognition performance with an extremely short bit stream, reducing the system's latency and energy consumption.
  • Compared with traditional stochastic computing networks, the network shortens the bit stream length by 64 times without affecting the recognition rate.
  • The research achieves a recognition rate of 99% with a 2-bit input, significantly reducing the system's latency and energy consumption.
  • The traditional 256-bit stochastic computing scheme is reduced by 82.87% in area and 1947 times in energy efficiency.
  • The proposed design has significant advantages in executing tasks such as image classification at the near-sensor end and edge devices.

Statistics:

  • The proposed network shortens the bit stream length by 64 times compared to traditional stochastic computing networks.
  • The network achieves a recognition rate of 99% with a 2-bit input.
  • The traditional 256-bit stochastic computing scheme is reduced by 82.87% in area and 1947 times in energy efficiency.
  • The proposed network reduces power consumption by 60.47% and area by 44.98% compared to traditional stochastic computing schemes.
  • The energy efficiency of the proposed network is increased by 12 times compared to traditional stochastic computing schemes.

Sources:

  • NewsRx. New Findings in Sensor Research Described from Beijing Information Science and Technology University (Design of Low-Cost and Highly Energy-Efficient Convolutional Neural Networks Based on Deterministic Encoding). Journal of Engineering. June 9, 2025; p 2101.
  • Design of Low-Cost and Highly Energy-Efficient Convolutional Neural Networks Based on Deterministic Encoding. Sensors, 2025, 25(10): 3127. (Sensors - http://www.mdpi.com/journal/sensors).
  • MDPI AG (publisher).

Read more