I-SPLIT: Deep Network Interpretability for Split Computing

被引:12
作者
Cunico, Federico [1 ]
Capogrosso, Luigi [1 ]
Setti, Francesco [1 ]
Carra, Damiano [1 ]
Fummi, Franco [1 ]
Cristani, Marco [1 ]
机构
[1] Univ Verona, Dept Comp Sci, Verona, Italy
来源
2022 26TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION (ICPR) | 2022年
关键词
D O I
10.1109/ICPR56361.2022.9956625
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
This work makes a substantial step in the field of split computing, i.e., how to split a deep neural network to host its early part on an embedded device and the rest on a server. So far, potential split locations have been identified exploiting uniquely architectural aspects, i.e., based on the layer sizes. Under this paradigm, the efficacy of the split in terms of accuracy can be evaluated only after having performed the split and retrained the entire pipeline, making an exhaustive evaluation of all the plausible splitting points prohibitive in terms of time. Here we show that not only the architecture of the layers does matter, but the importance of the neurons contained therein too. A neuron is important if its gradient with respect to the correct class decision is high. It follows that a split should be applied right after a layer with a high density of important neurons, in order to preserve the information flowing until then. Upon this idea, we propose Interpretable Split (I-SPLIT): a procedure that identifies the most suitable splitting points by providing a reliable prediction on how well this split will perform in terms of classification accuracy, beforehand of its effective implementation. As a further major contribution of I-SPLIT, we show that the best choice for the splitting point on a multiclass categorization problem depends also on which specific classes the network has to deal with. Exhaustive experiments have been carried out on two networks, VGG16 and ResNet-50, and three datasets, Tiny-Imagenet-200, notMNIST, and Chest X-Ray Pneumonia. The source code is available at https://github.com/vips4/I-Split.
引用
收藏
页码:2575 / 2581
页数:7
相关论文
共 36 条
[1]  
Adebayo J., 1992, ARXIV181003292
[2]  
[Anonymous], NOTMNIST DATASET
[3]  
Baehrens D, 2010, J MACH LEARN RES, V11, P1803
[4]   Machine Learning Interpretability: A Survey on Methods and Metrics [J].
Carvalho, Diogo, V ;
Pereira, Eduardo M. ;
Cardoso, Jaime S. .
ELECTRONICS, 2019, 8 (08)
[5]  
Choi H, 2018, IEEE IMAGE PROC, P3743, DOI 10.1109/ICIP.2018.8451100
[6]  
Cohen RobertA., 2020, 2020 IEEE International Conference on Multimedia and Expo (ICME), P1
[7]  
Du KX, 2019, 2019 IEEE 2ND 5G WORLD FORUM (5GWF), P135, DOI [10.1109/5gwf.2019.8911629, 10.1109/5GWF.2019.8911629]
[8]   BottleNet: A Deep Learning Architecture for Intelligent Mobile Cloud Computing Services [J].
Eshratifar, Amir Erfan ;
Esmaili, Amirhossein ;
Pedram, Massoud .
2019 IEEE/ACM INTERNATIONAL SYMPOSIUM ON LOW POWER ELECTRONICS AND DESIGN (ISLPED), 2019,
[9]   JointDNN: An Efficient Training and Inference Engine for Intelligent Mobile Cloud Computing Services [J].
Eshratifar, Amir Erfan ;
Abrishami, Mohammad Saeed ;
Pedram, Massoud .
IEEE TRANSACTIONS ON MOBILE COMPUTING, 2021, 20 (02) :565-576
[10]  
He Kaiming, 2016, Lecture Notes in Computer Science, V9908, P630, DOI [10.1007/978-3-319-46493-0_38, DOI 10.1007/978-3-319-46493-0_38, DOI 10.1109/CVPR.2016.90]