Neural Network Fixed-Point Control for Skipping Redundant Calculations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for reducing power consumption in object detection devices by dynamically controlling fixed-point positions in neural networks fail to optimize calculations effectively, leading to unnecessary power consumption due to continuous calculations on feature maps with similar values.
Innovation Solution
A data processing device and method that dynamically adjusts fixed-point positions in neural networks to skip calculations where the output is equivalent, increasing the equivalence rate and reducing power consumption by stopping clock cycles during redundant calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If dynamic fixed-point position control is implemented to reduce power consumption, then power consumption is reduced, but calculation accuracy may deteriorate due to increased equivalence rate
Solution Approach 1:
The patent dynamically adjusts fixed-point positions during calculation processing based on the actual data distribution and characteristics of each layer in the neural network. This allows the system to optimize the equivalence rate adaptively, skipping redundant calculations when possible while maintaining calculation accuracy when necessary, thus resolving the contradiction between power consumption reduction and accuracy maintenance
Solution Approach 2:
The patent changes the fixed-point position parameters dynamically during calculation based on statistical analysis of data distribution. By adjusting these parameters adaptively, the system can increase the equivalence rate to skip calculations (reducing power consumption) while ensuring that accuracy requirements are still met through selective application of the optimization
2Use of energy by stationary object
If data bit width is reduced to 8-16 bits for hardware implementation, then circuit scale and power consumption are reduced, but calculation precision is compromised
Solution Approach 1:
The patent applies different fixed-point positions to different layers and data types within the neural network calculation process. Instead of using a uniform bit width reduction approach, the system optimizes the precision locally for each layer based on its specific data distribution and requirements, thereby maintaining necessary calculation precision while still achieving overall power consumption reduction
3Reliability
If calculations are performed on all pixel values continuously, then calculation completeness is maintained, but power consumption increases due to redundant calculations on feature maps with similar values
Solution Approach 1:
The patent identifies and skips redundant calculations in feature maps where adjacent pixel values are equivalent or highly similar. By detecting these redundant regions and bypassing their calculation while maintaining overall calculation completeness through selective processing, the system significantly reduces power consumption without compromising the reliability of the final detection results
Data Source
Figure 1
Figure 2
Figure 3
AI summary
In a data processing device, a fixed-point position control unit determines, as first control. The fixed-point position control unit causes a detection calculation unit to perform calculation processing on processing target data at a processing point in time. The saturation rate control unit instructs, as second control to be repeated by the fixed-point position control unit, the fixed-point position control unit to move at least the fixed-point position as control to increase a lower limit saturation rate proportional to a magnitude of a counted lower limit counter value with respect to a result of the first control. The fixed-point position control unit performs, as the second control, a predetermined determination on the basis of the instruction from the saturation rate control unit and the metadata, determines the fixed-point position moved for each layer, and causes calculation processing to be performed.