CNN Operation Device Skipping Zero Pixels via Mosaic Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) consume significant power due to the extensive calculations required in their convolutional layers, leading to increased system load and energy consumption.
Innovation Solution
An operation device and method for a CNN that predicts zero outputs in the feature map, allowing for the skipping of operations on corresponding input pixels, utilizing a first operation unit for convolution based on a mosaic type, a prediction unit to identify skip pixels, and a second operation unit to process these pixels as zero, thereby reducing convolution operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolution operations are performed on all pixels of the input feature map, then the output feature map is accurately computed, but power consumption and system load increase significantly
Solution Approach 1:
The patent performs preliminary convolution operations only on selected pixels (e.g., every second pixel in a checkerboard pattern) before the final convolution step, rather than performing full convolutions on all pixels. This preliminary action on a subset of pixels allows prediction of zero outputs and skips unnecessary operations, reducing power consumption while maintaining output accuracy through the final consolidation step.
2Use of energy by moving object
If convolution operations are skipped for predicted zero pixels, then power consumption is reduced, but computational accuracy may be compromised
Solution Approach 1:
The patent uses feedback by comparing the preliminary convolution results with predicted patterns to identify pixels that will produce zero outputs. This feedback mechanism allows the system to confidently skip operations on these pixels without compromising accuracy, as the prediction is based on actual computed values from the preliminary step rather than random guessing.
Solution Approach 2:
The preliminary convolution step computes values for a subset of pixels that serve as the basis for predicting zero outputs. This preliminary action provides the information needed to make accurate predictions, ensuring that skipped operations truly result in zero values and thus maintaining overall output accuracy while reducing computational load.
3Reliability
If full convolution operations are performed, then complete feature map processing is achieved, but processing time increases
Solution Approach 1:
The patent performs a preliminary convolution step on a subset of pixels to gather information needed for predicting zero outputs. This preliminary action enables the system to identify and skip unnecessary full convolution operations, significantly reducing processing time while maintaining processing completeness through the final consolidation step that ensures all necessary features are captured.
4Use of energy by moving object
If prediction of zero pixels is implemented, then unnecessary operations are skipped, but device complexity increases
Solution Approach 1:
The patent segments the convolution process into distinct stages: a preliminary convolution step on selected pixels, a prediction step to identify zero outputs, and a final consolidation step. This segmentation allows the system to add prediction complexity only where necessary (on a subset of pixels) rather than throughout the entire feature map, managing device complexity while achieving energy savings.
Solution Approach 2:
The patent applies prediction logic locally only to specific pixels (e.g., every second pixel in a checkerboard pattern) rather than uniformly across the entire feature map. This local application of prediction reduces the overall complexity burden, as the prediction mechanism is instantiated only where needed to enable operation skipping, rather than being applied universally.
Data Source
AI summary
An operation device of a convolutional neural network includes a first operation unit performing a convolution operation on an input feature map based on a predetermined mosaic type to output at least one pixel information of an output feature map, a prediction unit predicting the other pixel information of the output feature map as one of a skip pixel and an operation pixel based on the at least one pixel information, and a second operation unit skipping the convolution operation with respect to the sip pixel to process the skip pixel as an output of zero.