MUXConv: Information Multiplexing in Convolutional Neural Networks

Convolutional neural networks have witnessed remarkable improvements in computational efficiency in recent years. A key driving force has been the idea of trading-off model expressivity and efficiency through a combination of $1\times 1$ and depth-wise separable convolutions in lieu of a standard convolutional layer. The price of the efficiency, however, is the sub-optimal flow of information across space and channels in the network. To overcome this limitation, we present MUXConv, a layer that is designed to increase the flow of information by progressively multiplexing channel and spatial information in the network, while mitigating computational complexity. Furthermore, to demonstrate the effectiveness of MUXConv, we integrate it within an efficient multi-objective evolutionary algorithm to search for the optimal model hyper-parameters while simultaneously optimizing accuracy, compactness, and computational efficiency. On ImageNet, the resulting models, dubbed MUXNets, match the performance (75.3% top-1 accuracy) and multiply-add operations (218M) of MobileNetV3 while being 1.6$\times$ more compact, and outperform other mobile models in all the three criteria. MUXNet also performs well under transfer learning and when adapted to object detection. On the ChestX-Ray 14 benchmark, its accuracy is comparable to the state-of-the-art while being $3.3\times$ more compact and $14\times$ more efficient. Similarly, detection on PASCAL VOC 2007 is 1.2% more accurate, 28% faster and 6% more compact compared to MobileNetV2. Code is available from https://github.com/human-analysis/MUXConv

PDF Abstract CVPR 2020 PDF CVPR 2020 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Semantic Segmentation ADE20K MUXNet-m + PPM Validation mIoU 35.8 # 144
Semantic Segmentation ADE20K MUXNet-m + C1 Validation mIoU 32.42 # 146
Pneumonia Detection ChestX-ray14 MUXNet-m AUROC 0.841 # 4
Params 2.1M # 3
FLOPS 200M # 4
Image Classification CIFAR-10 MUXNet-m Percentage correct 98.0 # 49
PARAMS 2.1M # 171
Neural Architecture Search CIFAR-10 MUXNet-m Top-1 Error Rate 2.0% # 7
Parameters 2.1M # 18
FLOPS 200M # 27
Image Classification CIFAR-100 MUXNet-m Percentage correct 86.1 # 54
PARAMS 2.1M # 156
Neural Architecture Search CIFAR-100 MUXNet-m FLOPS 200M # 7
Percentage Error 13.9 # 5
PARAMS 2.1M # 5
Neural Architecture Search CIFAR-10 Image Classification MUXNet-m Percentage error 2.0 # 5
Params 2.1M # 5
FLOPS 200M # 14
Image Classification ImageNet MUXNet-m Top 1 Accuracy 75.3% # 627
Top 5 Accuracy 92.5 # 221
Number of params 3.4M # 230
GFLOPs 0.436 # 43
Image Classification ImageNet MUXNet-xs Top 1 Accuracy 66.7% # 698
Top 5 Accuracy 86.8 # 254
Number of params 1.8M # 222
GFLOPs 0.132 # 5
Neural Architecture Search ImageNet MUXNet-xs Top-1 Error Rate 33.3 # 122
Accuracy 66.7 # 104
Params 1.8M # 53
MACs 66M # 62
Image Classification ImageNet MUXNet-s Top 1 Accuracy 71.6% # 675
Top 5 Accuracy 90.3 # 247
Number of params 2.4M # 223
GFLOPs 0.234 # 17
Neural Architecture Search ImageNet MUXNet-m Top-1 Error Rate 24.7 # 104
Accuracy 75.3 # 87
Params 3.4M # 51
MACs 218M # 66
Neural Architecture Search ImageNet MUXNet-l Top-1 Error Rate 23.4 # 75
Accuracy 76.6 # 63
Params 4.0M # 50
MACs 318M # 86
Image Classification ImageNet MUXNet-l Top 1 Accuracy 76.6% # 590
Top 5 Accuracy 93.2 # 202
Number of params 4.0M # 235
Hardware Burden None # 1
Operations per network pass None # 1
GFLOPs 0.636 # 67
Neural Architecture Search ImageNet MUXNet-s Top-1 Error Rate 28.4 # 119
Accuracy 71.6 # 101
Params 2.4M # 52
MACs 117M # 63

Methods


No methods listed for this paper. Add relevant methods here