Fixed-Point Circuitry for CNN Layer Overflow Prevention
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for fixed-point processing in CNNs often lead to overflow issues, resulting in reduced calculation accuracy due to the need for uniform processing across layers with significantly different value ranges.
Innovation Solution
An image processing controller with a fixed-point circuitry that determines a design fixed-point range for each neural network block based on its specific maximum and minimum values, performing fixed-point processing within this range to prevent overflow and maintain accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If uniform fixed-point processing is applied across all CNN layers, then the processing is simple and fast, but overflow occurs and calculation accuracy deteriorates
Solution Approach 1:
The patent applies different fixed-point processing parameters to different CNN layers based on their specific characteristics. Each layer has its own maximum and minimum value ranges, and the fixed-point processing is customized for each layer rather than using a uniform approach across all layers. This resolves the contradiction by maintaining processing efficiency while improving accuracy through localized optimization.
Solution Approach 2:
The patent dynamically adjusts fixed-point processing parameters (such as scaling factors and offset values) based on the actual data ranges observed in each CNN layer. By changing these parameters adaptively rather than using fixed uniform values, the system prevents overflow in layers with smaller value ranges while maintaining efficiency. This resolves the contradiction between uniform processing simplicity and accuracy requirements.
2Speed
If fixed-point conversion is performed for hardware deployment, then real-time processing capability is improved, but overflow problems occur leading to accuracy loss
Solution Approach 1:
The patent performs preliminary analysis of the data ranges (maximum and minimum values) for each CNN layer before deploying fixed-point processing. By pre-calculating the appropriate fixed-point parameters based on the actual data characteristics of each layer, the system ensures that the fixed-point conversion will not cause overflow while maintaining real-time processing capability. This preliminary preparation resolves the contradiction between speed and accuracy.
Data Source
AI summary
An image processing controller for a display device includes a fixed-point circuitry and a plurality of neural network blocks cascaded sequentially. The fixed-point circuitry is electrically connected to each neural network block, and configured to receive a feature signal corresponding to an output feature map about the display device from each neural network block, perform fixed-point processing on the feature signal to acquire fixed-point data within a design accuracy range, and input the fixed-point data as an input feature map to a next-level neural network block.