Superpixel CNN Layers for Efficient Shallow-Depth Computation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing convolutional neural networks (CNNs) exhibit poor computational efficiency when input or output depths are shallow, leading to inefficient matrix multiplication operations, particularly in software libraries and hardware units.
Innovation Solution
The implementation of superpixel CNN layers, which group multiple inputs together, trading spatial extent for depth extent, allowing for more efficient matrix multiplication operations and reducing the number of kernel elements required, thereby enhancing computational efficiency and usability without modifying hardware.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If standard CNN layers are used with shallow input or output depths, then the network structure remains simple, but computational efficiency deteriorates due to inefficient matrix multiplication operations
Solution Approach 1:
The patent applies dimensionality change by transforming the input tensor from spatial dimensions (X×Y) to a combined depth dimension (Z'). Specifically, it reshapes the input tensor of shape (X, Y, Z) into (X', Y', Z') where Z' = Z × N × M and X' = X/(N×M), effectively moving the problem from spatial convolution to depth-wise operations that are more efficient for matrix multiplication on modern hardware
Solution Approach 2:
The patent changes the parameters of the convolutional layer by introducing a transformation matrix T that modifies the weight matrices. The transformed weight matrices operate on the transformed input tensor to produce transformed output tensors, which are then reshaped to match the original output format. This parameter transformation enables efficient computation while maintaining mathematical equivalence
2Productivity
If superpixel CNN layers are implemented to improve computational efficiency, then matrix multiplication becomes more efficient, but the number of kernel elements increases
Solution Approach 1:
The patent merges multiple kernel elements into a single transformed weight matrix that operates on superpixel inputs. By grouping N×M input pixels into superpixels and transforming the weight matrices accordingly, it reduces the number of separate kernel operations needed while maintaining the same computational functionality through more efficient matrix multiplication
3Speed
If standard CNN layers are used, then hardware compatibility is good, but computational latency increases due to inefficient operations
Solution Approach 1:
The patent substitutes the traditional spatial convolution mechanism with a transformed matrix multiplication mechanism. By transforming the input tensor and weight matrices into a format that leverages modern hardware's matrix multiplication capabilities, it replaces inefficient spatial operations with optimized linear algebra operations that execute faster on current computing architectures
Data Source
AI summary
Methods, systems, and apparatus for efficiently performing a computation of a convolutional neural network layer. One of the methods includes transforming a X by Y by Z input tensor into a X′ by Y′ by Z′ input tensor, wherein X′ is smaller than or equal to X, Y′ is smaller than or equal to Y, and Z′ is larger than or equal to Z; obtaining one or more modified weight matrices, wherein the modified weight matrices operate on the X′ by Y′ by Z′ input tensor to generate a U′ by V′ by W′ output tensor, and the U′ by V′ by W′ output tensor is a transformed U by V by W output tensor; and processing the X′ by Y′ by Z′ input tensor using the modified weight matrices to generate the U′ by V′ by W′ output tensor.