IP Library › Granted Patent US 12,488,248
Granted Patent B2
US 12,488,248 · App. 17/174,049 · Granted Dec 2, 2025

Deep neural network compression based on filter importance

Inventors: Rajesh Kumar Tamada (Bengaluru, IN); Junpyo Hong (San Ramon, CA); Attila Márk Rádics (Nagykovácsi, HU); Hans Krupakar (San Ramon, CA); Venkata Ratnam Saripalli (Danville, CA); Dibyajyoti Pati (Dublin, CA); Guarav Kumar (Bokaro Steel, IN)
Assignee: GE Precision Healthcare LLC
G06N3/082G06F18/2115G06F18/214G06N3/063G06N5/04
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,488,248
App. No.
17/174,049
Granted
Dec 2, 2025
Kind
B2
Abstract

Techniques are provided for compressing deep neural networks using a structured filter pruning method that is extensible and effective. According to an embodiment, a computer-implemented method comprises determining, by a system operatively coupled to a processor, importance scores for filters of layers of a neural network model previously trained until convergence for an inferencing task on a training dataset. The method further comprises removing, by the system, a subset of the filters from one or more layers of the layers based on the importance scores associated with the subset failing to satisfy a threshold importance score value. The method further comprises converting, by the system, the neural network model into a compressed neural network model with the subset of the filters removed.

Claims (75)

1 . A system, comprising:

a memory; and

a processor that executes computer executable instructions stored in the memory to perform operations comprising:

training, using a training dataset, until a defined convergence criterion is satisfied, a neural network to perform a medical image inferencing task associated with medical images of patients, wherein the medical image inferencing task comprises modifications to the medical images;

determining a baseline score for a performance metric for the neural network based on the training;

determining a threshold importance criterion based on a defined criterion with respect to the baseline score for the performance metric;

compressing the neural network into a compressed neural network that employs less computing hardware resources for execution than the neural network while satisfying the defined criterion with respect to the baseline score for the performance metric,

wherein the compressing comprises:

determining, based on a single forward pass of the training dataset through the neural network, respective importance scores for filters of layers of the neural network with respect to the performance metric, wherein the determining the respective importance scores for the filters comprises, for each filter:

determining respective weight scores for the filter to each data point of the training dataset, and

determining the importance score for the filter based on an aggregation of the respective weight scores for the filter to the data points of the training dataset;

removing a subset of the filters from the layers of the neural network having respective importance scores that satisfy the threshold importance criterion, resulting in the compressed neural network, wherein the layers of the neural network comprise an input layer, at least one convolution layer, and an output layer, and wherein the threshold importance criterion comprises:

a first threshold importance score for determining whether to remove one or more first filters of at least one of the input layer or the output layer, and

a second threshold importance score employed for determining whether to remove one or more second filters of the at least one convolution layer, wherein the first threshold importance score is lower than the second threshold importance score; and

training, using the training dataset, until the defined convergence criterion is satisfied, the compressed neural network to perform the medical image inferencing task; and

performing, using the compressed neural network, the medical image inferencing task with a new dataset to generate a medical inference output.

2 . The system of claim 1 , wherein the determining the respective importance scores for the filters further comprises employing a gradient-based scoring mechanism.

3 . The system of claim 1 , wherein the determining the respective weight score for the filter to each data point of the training dataset comprises determining sensitivity gradients of weights of the filter without changing the weights of the filter.

4 . The system of claim 1 , wherein the determining the respective importance scores for each of the filters occurs independently without considering their impact on other layers.

5 . The system of claim 1 , wherein the removing the subset of the filters further comprises ranking the filters within each layer of the layers relative to one another as a function of their importance scores.

6 . The system of claim 1 , wherein the layers comprise at least one convolutional layer that precedes at least one batch normalization layer, and wherein based on the subset of the filters including a first filter included in the at least one convolutional layer, the removing further removes a second filter included in the at least one batch normalization layer that corresponds to the first filter.

7 . The system of claim 1 , wherein the filters comprise residual connection filters shared between two or more layers of the layers, and wherein the removing applies an exception for removing the residual connection filters regardless of whether their importance scores fail to satisfy the threshold importance criterion.

8 . The system of claim 1 , wherein the neural network comprises at least two layers with residual connections between two or more filters of the filters, and wherein the removing maintains the two or more filters regardless of whether their importance scores fail to satisfy the threshold importance criterion.

9 . The system of claim 1 , wherein the determining the respective importance scores for the filters further comprises:

normalizing the respective importance scores of the filters within each layer of the layers, resulting in respective normalized importance scores for the filters within each layer,

determining cumulative sums of the respective normalized importance scores for each layer, and

sorting the layers as a function of the cumulative sums associated with each layer.

10 . The system of claim 9 , wherein the determining the threshold importance criterion is further based on a measure of contribution of the respective normalized importance scores to the cumulative sums associated with each layer.

11 . The system of claim 1 , wherein the compressed neural network has a smaller memory footprint and higher inferencing speed relative to the neural network.

12 . The system of claim 1 , wherein the compressed neural network has less than or equal to a 15% performance accuracy degradation relative to the neural network.

13 . The system of claim 12 , wherein the medical image inferencing task further comprises at least one of an image segmentation task or an image classification task.

14 . The system of claim 1 , wherein first filters having respective importance scores below the first threshold importance score are removed, and wherein second filters having respective importance scores under the second threshold importance score are removed.

15 . A method comprising:

training, by a system operatively coupled to a processor, using a training dataset, until a defined convergence criterion is satisfied, a neural network to perform a medical image inferencing task associated with medical images of patients, wherein the medical image inferencing task comprises modifications to the medical images;

determining, by the system, a baseline score for a performance metric for the neural network based on the training;

determining, by the system, a threshold importance criterion based on a defined criterion with respect to the baseline score for the performance metric;

compressing, by the system, the neural network into a compressed neural network that employs less computing hardware resources for execution than the neural network while satisfying the defined criterion with respect to the baseline score for the performance metric, wherein the compressing comprises:

determining, based on a single forward pass of the training dataset through the neural network, respective importance scores for filters of layers of the neural network with respect to the performance metric, wherein the determining the respective importance scores for the filters comprises, for each filter:

determining respective weight scores for the filter to each data point of the training dataset, and

determining the importance score for the filter based on an aggregation of the respective weight scores for the filter to the data points of the training dataset;

removing a subset of the filters from the layers of the neural network having respective importance scores that satisfy the threshold importance criterion, resulting in the compressed neural network, wherein the layers of the neural network comprise an input layer, at least one convolution layer, and an output layer, and wherein the threshold importance criterion comprises:

a first threshold importance score for determining whether to remove one or more first filters of at least one of the input layer or the output layer, and

a second threshold importance score employed for determining whether to remove one or more second filters of the at least one convolution layer, wherein the first threshold importance score is lower than the second threshold importance score; and

training, using the training dataset, until the defined convergence criterion is satisfied, the compressed neural network to perform the medical image inferencing task; and

performing, by the system, using the compressed neural network, the medical image inferencing task with a new dataset to generate a medical inference output.

16 . The method of claim 15 , wherein the determining the respective importance scores comprises employing a gradient-based scoring mechanism.

17 . The method of claim 15 , wherein the determining the respective importance scores further comprises determining the respective importance scores each of the filters independently without considering their impact on other layers.

18 . The method of claim 15 , wherein the removing the subset of the filters further comprises:

ranking the filters within each layer of the layers relative to one another as a function of their importance scores.

19 . The method of claim 15 , wherein the layers comprise at least one convolutional layer that precedes at least one batch normalization layer, and wherein based on the subset of the filters comprising a first filter included in the at least one convolutional layer, the compressing the neural network further comprises:

removing a second filter included in the at least one batch normalization layer that corresponds to the first filter.

20 . The method of claim 15 , wherein the filters comprise residual connection filters shared between two or more layers of the layers, and wherein the removing comprises applying an exception for removing the residual connection filters regardless of whether their importance scores fail to satisfy the threshold importance criterion.

21 . The method of claim 15 , determining the respective importance scores for the filters further comprising:

normalizing the respective importance scores of the filters within each layer of the layers, resulting in respective normalized importance scores for the filters within each layer;

determining cumulative sums of the respective normalized importance scores for each layer; and

sorting the layers as a function of the cumulative sums associated with each layer.

22 . The method of claim 21 , wherein the determining the threshold importance criterion is further based on a measure of contribution of the respective normalized importance scores to the cumulative sums associated with each layer.

23 . The method of claim 15 , wherein the medical image inferencing task comprises at least one of an image segmentation task or an image classification task.

24 . The method of claim 15 , wherein the compressed neural network has a smaller memory footprint and higher inferencing speed relative to the neural network.

25 . The method of claim 15 , wherein first filters having respective importance scores below the first threshold importance score are removed, and wherein second filters having respective importance scores under the second threshold importance score are removed.

26 . A non-transitory machine-readable medium, comprising executable instructions that, when executed by a processor, facilitate performance of operations, comprising:

training, using a training dataset, until a defined convergence criterion is satisfied, a neural network to perform a medical image inferencing task associated with medical images of patients, wherein the medical image inferencing task comprises modifications to the medical images;

determining a baseline score for a performance metric for the neural network model based on the training;

determining a threshold importance criterion based on a defined criterion with respect to the baseline score for the performance metric;

compressing the neural network into a compressed neural network that employs less computing hardware resources for execution than the neural network while satisfying the defined criterion with respect to the baseline score for the performance metric, wherein the compressing comprises:

determining, based on a single forward pass of the training dataset through the neural network, respective importance scores for filters of layers of the neural network with respect to the performance metric, wherein the determining the respective importance scores for the filters comprises, for each filter:

determining respective weight scores for the filter to each data point of the training dataset, and

determining the importance score for the filter based on an aggregation of the respective weight scores for the filter to the data points of the training dataset;

removing a subset of the filters from the layers of the neural network having respective importance scores that satisfy the threshold importance criterion, resulting in the compressed neural network, wherein the layers of the neural network comprise an input layer, at least one convolution layer, and an output layer, and wherein the threshold importance criterion comprises:

a first threshold importance score for determining whether to remove one or more first filters of at least one of the input layer or the output layer, and

a second threshold importance score employed for determining whether to remove one or more second filters of the at least one convolution layer, wherein the first threshold importance score is lower than the second threshold importance score,

training, using the training dataset, until the defined convergence criterion is satisfied, the compressed neural network to perform the medical image inferencing task; and

performing, using the compressed neural network, the medical image inferencing task with a new dataset to generate a medical inference output.

27 . The non-transitory machine-readable medium of claim 26 , wherein the determining the respective weight scores for the filters to each data point of the training dataset comprises determining sensitivity gradients of weights of the filter without changing the weights of the filter.

28 . The non-transitory machine-readable medium of claim 26 , wherein first filters having respective importance scores below the first threshold importance score are removed, and wherein second filters having respective importance scores under the second threshold importance score are removed.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Feb 11, 2021
From: KUMAR, TAMADA RAJESH; HONG, JUNPYO; RÁDICS, ATTILA MÁRK; KRUPAKAR, HANS; SARIPALLI, VENKATA RATNAM; PATI, DIBYAJYOTI; KUMAR, GUARAV
To: GE PRECISION HEALTHCARE LLC
Reel/Frame 055237/0354 →
Continuity (1)
Related Publication 20220253708A1 · Aug 11, 2022
References Cited (62)
US 20190050715A1 · Ooi · 2019 [cited by examiner]
CN 106599883A · 2017 [cited by examiner]
WO WO2020033898A1 · 2020 [cited by examiner]
Z. Liu, Z. Chen and W. Li, “Importance-Aware Filter Selection for Convolutional Neural Network Acceleration,” 2019 IEEE Visual Communications and Image Processing (VCIP), Sydney, NSW, Australia, 2019, pp. 1-4 (Year: 201… [cited by examiner]
Li, Hao, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. “Pruning filters for efficient convnets.” arXiv preprint arXiv:1608.08710 (2016). (Year: 2016). [cited by examiner]
Jingjing Cao, Sam Kwong, Ran Wang, Xiaodong Li, Ke Li, Xiangfei Kong, Class-specific soft voting based multiple extreme learning machines ensemble, Neurocomputing, vol. 149, Part A, (Year: 2015). [cited by examiner]
A. Salama, O. Ostapenko, T. Klein and M. Nabi, “Prune Your Neurons Blindly: Neural Network Compression through Structured Class-blind Pruning,” ICASSP 2019—2019 IEEE International Conference on Acoustics, Speech and Sig… [cited by examiner]
Xu, Xiaowei, Qing Lu, Lin Yang, Sharon Hu, Danny Chen, Yu Hu, and Yiyu Shi. “Quantization of fully convolutional networks for accurate biomedical image segmentation.” In Proceedings of the IEEE conference on computer vi… [cited by examiner]
Frankle, et al., “The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks,” arXiv:1803.03635v5 [cs.LG] Mar. 4, 2019, https://arxiv.org/abs/1803.03635, 42 pages. [cited by applicant]
Ding, et al., “Global Sparse Momentum SGD for Pruning VeryDeep Neural Networks,” arXiv:1909.12778v3 [cs.LG] Oct. 25, 2019, https://arxiv.org/pdf/1909.12778.pdf, 13 pages. [cited by applicant]
github.com, “he-y / Awesome-Pruning,” https://github.com/he-y/Awesome-Pruning, 10 pages. [cited by applicant]
Li, et al., “PENNI: Pruned Kernel Sharing for Efficient CNN Inference,” arXiv:2005.07133v2 [cs.CV] Jun. 25, 2020, https://arxiv.org/pdf/2005.07133.pdf, 15 pages. [cited by applicant]
Wang, et al., “Accelerate Your CNN from Three Dimensions: A Comprehensive Pruning Framework,” arXiv:2010.04879v1 [cs.CV] Oct. 10, 2020, https://arxiv.org/pdf/2010.04879v1.pdf, 8 pages. [cited by applicant]
Su, et al., “Sanity-Checking Pruning Methods: Random Tickets can Win the Jackpot,” arXiv:2009.11094v1 [cs.LG] Sep. 22, 2020, https://arxiv.org/pdf/2009.11094v1.pdf, 14 pages. [cited by applicant]
Wang, et al., “Achieving Adversarial Robustness via Sparsity,” arXiv:2009.05423v1 [cs.LG] Sep. 11, 2020, https://arxiv.org/pdf/2009.05423v1.pdf, 9 pages. [cited by applicant]
Paganini, “Prune Responsibly,” arXiv:2009.09936v1 [cs.CV] Sep. 10, 2020, https://arxiv.org/pdf/2009.09936v1.pdf, 18 pages. [cited by applicant]
Wang, et al., “SparseRT: Accelerating Unstructured Sparsity on GPUs for DeepLearning Inference,” arXiv:2008.11849v1 [cs.LG] Aug. 26, 2020, https://arxiv.org/pdf/2008.11849v1.pdf, 12 pages. [cited by applicant]
Rastegari, et al., “XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks,” arXiv:1603.05279v4 [cs.CV] Aug. 2, 2016, https://arxiv.org/abs/1603.05279, 17 pages. [cited by applicant]
Zhou, et al., “Incremental Network Quantization: Towards Lossless CNNs with Low-Precision Weights,” arXiv:1702.03044v2 [cs.CV] Aug. 25, 2017, https://arxiv.org/abs/1702.03044, 14 pages. [cited by applicant]
Zhuang, et al., “Towards Effective Low-bitwidth Convolutional Neural Networks,” arXiv:1711.00205v2 [cs.CV] Nov. 17, 2017, https://arxiv.org/abs/1711.00205, 11 pages. [cited by applicant]
Guo, et al., “Dynamic Network Surgery for Efficient DNNs,” arXiv:1608.04493v2 [cs.NE] Nov. 10, 2016, https://arxiv.org/abs/1608.04493, 9 pages. [cited by applicant]
Han, et al., “Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding,” arXiv:1510.00149v5 [cs.CV] Feb. 15, 2016, https://arxiv.org/abs/1510.00149?source=post_page—, 5 pa… [cited by applicant]
Alvarez, et al., “Learning the Number of Neurons in Deep Networks,” arXiv:1611.06321v3 [cs.CV] Oct. 11, 2018, https://arxiv.org/abs/1611.06321, 9 pages. [cited by applicant]
Denton, et al., “Exploiting Linear Structure Within Convolutional Networks for Efficient Evaluation,” arXiv:1404.0736v2 [cs.CV] Jun. 9, 2014, https://arxiv.org/abs/1404.0736, 11 pages. [cited by applicant]
Le Cun, et al., “Optimal brain damage,” Advances in neural information processing systems, published Jan. 1, 1989, 8 pages. [cited by applicant]
Hassibi, et al., “Second order derivatives for network pruning: Optimal Brain Surgeon,” Part of Advances in Neural Information Processing Systems 5 (NIPS 1992), 8 pages. [cited by applicant]
Li, et al., “Pruning Filters for Efficient ConvNets,” arXiv:1608.08710v3 [cs.CV] Mar. 10, 2017, https://arxiv.org/abs/1608.08710, 13 pages. [cited by applicant]
Molchanov, et al., “Pruning Convolutional Neural Networks for Resource Efficient Inference,” harXiv:1611.06440v2 [cs.LG] Jun. 8, 2017, ttps://arxiv.org/abs/1611.06440, 17 pages. [cited by applicant]
Krizhevsky, “The CIFAR-10 dataset,” 2009, 4 pages. [cited by applicant]
nih.gov, “Montgomery County—Chest X-ray Database,” openi.nlm.nih.gov/imgs/collections/NLM-MontgomeryCXRSet.zip, 2 pages, Retrieved from the Internet: Dec. 10, 2020. [cited by applicant]
Kingma, et al., “Adam: A Method for Stochastic Optimization,” arXiv:1412.6980v9 [cs.LG] Jan. 30, 2017, https://arxiv.org/abs/1412.6980, 15 pages. [cited by applicant]
Ronneberger, et al., U-Net: Convolutional Networks for Biomedical Image Segmentation, arXiv:1505.04597v1 [cs.CV] May 18, 2015, https://arxiv.org/abs/1505.04597, 8 pages. [cited by applicant]
Simonyan, et al., “Very Deep Convolutional Networks for Large-Scale Image Recognition,” arXiv:1409.1556v6 [cs.CV] Apr. 10, 2015, https://arxiv.org/abs/1409.1556, 14 pages. [cited by applicant]
He, et al., “Deep Residual Learning for Image Recognition,” harXiv:1512.03385v1 [cs.CV] Dec. 10, 2015, https://arxiv.org/abs/1512.03385, 12 pages. [cited by applicant]
Park, et al., “Lookahead: A Far-sighted Alternative of Magnitude-based Pruning,” Under review as a conference paper at ICLR 2020, Retrieved from the Internet: Dec. 10, 2020, https://openreview.net/attachment?d=ryl3ygHYD… [cited by applicant]
Li, et al., “PENNI: Pruned Kernel Sharing for Efficient CNN Inference,” arXiv:2005.07133v2 [cs.CV] Jun. 25, 202, https://arxiv.org/abs/2005.07133, 15 pages. [cited by applicant]
Nachar, “The Mann-Whitney U: A Test for Assessing Whether Two Independent Samples Come from the Same Distribution,” Tutorials in Quantitative Methods for Psychology 2008, vol. 4(1), p. 13-20. [cited by applicant]
cs.toronto.edu, “The CIFAR-10 dataset,” Alex Krizhevsky, 2009, https://www.cs.toronto.edu/%7Ekriz/cifar.html, 4 pages. [cited by applicant]
openi.nlm.nih.gov, “Montgomery County Chest X-ray Database,” openi.nlm.nih.gov/imgs/collections/NLM-MontgomeryCXRSet.zip, 2 pages. [cited by applicant]
Bellec, et al., “Deep Rewiring: Training very sparse deep networks,” arXiv:1711.05136v5 [cs.NE] Aug. 7, 2018, https://arxiv.org/pdf/1711.05136.pdf, 24 pages. [cited by applicant]
Dai, et al., “Compressing Neural Networks using the Variational Information Bottleneck,” arXiv:1802.10399v3 [cs.CV] Apr. 19, 2018, https://arxiv.org/pdf/1802.10399.pdf, 27 pages. [cited by applicant]
Deng, et al., “ImageNet: a Large-Scale Hierarchical Image Database,” Conference: 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), Jun. 20-25, 2009, Miami, Florida, USA, https… [cited by applicant]
Ding, et al., “Centripetal SGD for Pruning Very Deep Convolutional Networks with Complicated Structure,” arXiv:1904.03837v1 [cs.LG] Apr. 8, 2019, https:/arxiv.org/pdf/1904.03837.pdf, 11 pages. [cited by applicant]
Ding, et al., “Global Sparse Momentum SGD for Pruning Very Deep Neural Networks,” rXiv:1909.12778v3 [cs.LG] Oct. 25, 2019 ,https://arxiv.org/pdf/1909.12778.pdf, 13 pages. [cited by applicant]
Frankle, et al., “The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks,” arXiv:1803.03635v5 [cs.LG] Mar. 4, 2019, https://arxiv.org/pdf/1803.03635.pdf, 42 pages. [cited by applicant]
Hassibi, et al., “Second order derivatives for network pruning: Optimal brain surgeon,” Advances in Neural Information Processing Systems 5 (NIPS 1992), https://authors.library.caltech.edu/54983/3/647-second-order-deriv… [cited by applicant]
He, et al., “Deep Residual Learning for Image Recognition,” arXiv:1512.03385v1 [cs.CV] Dec. 10, 2015, https://arxiv.org/pdf/1512.03385.pdf, 12 pages. [cited by applicant]
He, et al., “AMC: AutoML for Model Compression and Acceleration on Mobile Devices,” arXiv:1802.03494v4 [cs.CV] Jan. 16, 2019, https://arxiv.org/pdf/1802.03494.pdf, 17 pages. [cited by applicant]
Kingma, et al., “Adam: A Method for Stochastic Optimization,” rXiv:1412.6980v9 [cs.LG] Jan. 30, 2017, https://arxiv.org/pdf/1412.6980.pdf, 15 pages. [cited by applicant]
Lecun, et al., “Optimal Brain Damage,” Conference: Advances in Neural Information Processing Systems 2, [NIPS Conference, Denver, Colorado, USA, Nov. 27-30, 1989], https://www.researchgate.net/publication/221618539_Opti… [cited by applicant]
Lin, et al., “Runtime Neural Pruning,” 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA, https://dl.acm.org/doi/pdf/10.5555/3294771.3294979, 11 pages. [cited by applicant]
Liu, et al., “Learning Efficient Convolutional Networks through Network Slimming,” arXiv:1708.06519v1 [cs.CV] Aug. 22, 2017, https://arxiv.org/pdf/1708.06519.pdf, 10 pages. [cited by applicant]
Louizos, et al., “Bayesian Compression for Deep Learning,” arXiv:1705.08665v4 [stat.ML] Nov. 6, 2017, https://arxiv.org/pdf/1705.08665.pdf, 17 pages. [cited by applicant]
Louizos, et al., “Learning Sparse Neural Networks through L0 Regularization,” arXiv:1712.01312v2 [stat.ML] Jun. 22, 2018, https://arxiv.org/pdf/1712.01312.pdf 13 pages. [cited by applicant]
Molchanov, et al., “Variational Dropout Sparsifies Deep Neural Networks,” aarXiv:1701.05369v3 [stat.ML] Jun. 13, 2017, https://arxiv.org/pdf/1701.05369.pdf, 10 pages. [cited by applicant]
Neal, “Bayesian learning for neural networks,” vol. 118. Springer Science & Business Media, 2012, https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.446.9306&rep=rep1&type=pdf, 195 pages. [cited by applicant]
Park, et al., “Lookahead: A Far-Sightedalternative Ofmagnitude-Basedpruning,” Published as a conference paper at ICLR Apr. 26, 2020, https://openreview.net/pdf?id=ryl3ygHYDB, 20 pages. [cited by applicant]
Ronnebeger et al., “U-Net: Convolutional Networks for Biomedical Image Segmentation,” arXiv:1505.04597v1 [cs.CV] May 18, 2015, https://arxiv.org/pdf/1505.04597.pdf, 8 pages. [cited by applicant]
Wang, “Pruning Convolutional Filters with First OrderTaylor Series Ranking,” ABCs, 2018, http://users.cecs.anu.edu.au/˜Tom.Gedeon/conf/ABCs2018/paper/ABCs2018_paper_53.pdf, 7 pages. [cited by applicant]
Wen, “Learning Structured Sparsity in Deep Neural Networks,” arXiv:1608.03665v4 [cs.NE] Oct. 18, 2016, https://arxiv.org/pdf/1608.03665.pdf, 10 pages. [cited by applicant]
Zhang, et al., “Learning to Share: Simultaneous Parameter Tying and Sparsification in Deep Learning,” Published as a conference paper at ICLR 2018, https://openreview.net/pdf?id=rypT3fb0b, 14 pages. [cited by applicant]
Wilcoxon et al., “Critical Values and Probability Levels for the Wilcoxon Rank Sum Test and the Wilcoxon Signed Rank Test,” Selected Tables in Mathematical Statistics, vol. 1, 1970, pp. 171-259. [cited by applicant]