Neural Network Blocks Using Grouped Convolution for Mobile Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network architectures for mobile devices, such as MobileNets, face inefficiencies in computational burden and accuracy when implemented on constrained devices like smartphones, despite using techniques like hard swish activation functions and inverted residual blocks.
Innovation Solution
A neural network block comprising an expansion layer, a compression layer, and a grouped convolution layer, with normalization techniques applied during the inference phase using calculated scaling and offset values based on standard deviations, replacing depth-wise convolution with grouped convolution, and using fixed-point arithmetic for weight storage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If depth-wise convolution is used in MobileNet blocks, then computational burden is reduced, but classification accuracy deteriorates
Solution Approach 1:
The patent segments the convolution operation into grouped convolution, dividing channels into groups and applying separate convolutions to each group. This segmentation allows the network to capture more diverse feature relationships compared to depth-wise convolution, thereby improving classification accuracy while maintaining computational efficiency through the grouped structure.
Solution Approach 2:
The patent changes the convolution parameter from depth-wise (1x1 kernel per channel) to grouped convolution (kxk kernels across channel groups). This parameter change enables better spatial feature extraction while controlling computational cost through the grouping mechanism, resolving the accuracy-efficiency tradeoff.
2Speed
If batch normalisation is applied after each layer, then training speed is improved, but inference phase computational overhead increases
Solution Approach 1:
The patent applies batch normalisation during the training phase to pre-normalize the data distributions and stabilize training. During inference, the pre-computed running mean and variance from training are used, eliminating the need for real-time normalization calculations and reducing inference complexity while maintaining training efficiency.
Solution Approach 2:
The patent copies the normalization parameters (running mean and variance) computed during training to the inference phase. This copying allows the inference phase to use fixed parameters without performing complex normalization operations, reducing computational overhead while preserving the benefits of normalization trained during the training phase.
3Measurement precision
If floating-point arithmetic is used for weight storage, then precision is improved, but memory usage and computational burden increase
Solution Approach 1:
The patent changes the numerical precision parameter from floating-point to fixed-point arithmetic for weight storage and computation. This parameter change reduces memory usage and computational burden by using simpler integer-based arithmetic, while carefully selected fixed-point precision maintains sufficient accuracy for mobile device classification tasks.
Solution Approach 2:
The patent uses fixed-point arithmetic as a simpler, cheaper alternative to floating-point arithmetic for weight storage and computation. This approach sacrifices some precision but gains significant reductions in memory usage and computational complexity, making the neural network more suitable for resource-constrained mobile devices.
Data Source
AI summary
A neural network block includes a plurality of layers arranged sequentially. Each layer includes an expansion layer having a first number of input channels and a second number of output channels, where the second number is larger than the first number, a compression layer, having a third number of input channels and a fourth number of output channels, wherein the fourth number is smaller than the third number, and a grouped convolution layer.