Spatial Adaptive Separable Convolutional Layer for Neural Network Feature Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The complexity of convolutional neural networks increases significantly with the number of features to be identified in input data, leading to increased resource and computational requirements due to the need for multiple filters in each convolutional layer, especially when dealing with multi-layered input data.
Innovation Solution
The implementation of a spatial adaptive separable convolutional layer that applies multiple filters of varying sizes to different layers, allowing for parallel processing and reducing the computational load by using depthwise and pointwise convolutions, thereby generating compact feature descriptors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple filters are applied to each layer in a set of input data items to identify features, then feature identification capability is improved, but computational complexity and resource requirements increase greatly
Solution Approach 1:
The patent segments the convolutional layer into two distinct operations: depthwise convolution that applies a single filter to each input channel independently, and pointwise convolution that uses 1x1 convolutions to combine features across channels. This segmentation reduces the number of filter applications needed while maintaining feature extraction capability, directly addressing the contradiction between feature identification capability and computational complexity
Solution Approach 2:
The patent introduces an intermediary pointwise convolution operation between the depthwise convolution and the final feature map generation. This intermediary step efficiently aggregates features from multiple depthwise convolution outputs using 1x1 convolutions, reducing computational complexity while preserving the ability to identify multiple features across different input layers
2Measurement precision
If multiple filters are applied to each layer in a set of input data items to identify features, then feature identification capability is improved, but memory and computing resources increase
Solution Approach 1:
By segmenting the convolutional operation into depthwise and pointwise components, the patent reduces the total number of filters needed. Each input channel is processed by a single depthwise filter, and feature aggregation is achieved through efficient pointwise convolutions, significantly reducing memory requirements compared to applying multiple full convolutional filters to each input layer
Solution Approach 2:
The patent changes the parameters of the convolutional operation by using 1x1 convolution kernels in the pointwise operation. This parameter change reduces the computational burden and memory requirements while maintaining the ability to combine and aggregate features effectively, directly addressing the resource consumption issue
3Measurement precision
If individual filters are applied to each layer to produce individual output, then feature extraction accuracy is improved, but processing time increases
Solution Approach 1:
The patent segments the feature extraction process into parallel depthwise convolution operations that can be executed independently for each input channel, followed by a pointwise convolution that aggregates the results. This segmentation enables parallel processing of individual channels, reducing overall processing time while maintaining feature extraction accuracy through the subsequent aggregation step
Data Source
AI summary
Apparatuses, systems, and techniques to identify features within one or more images. Features are identified in one or more images using one or more neural networks containing convolutional layers with multiple filters that may be executed by one or more parallel processing unit.