SlideShare a Scribd company logo
Copyright © 2017 Almond Research Pte Ltd. 1
Matthew Chiu
May 2nd 2017
Designing CNN Algorithms For Real-
time Applications
Copyright © 2017 Almond Research Pte Ltd. 2
• Consider 3 stages: Dataset – Model – Deployment
• Understand relationship between size (depth) and
performance of Deep Neural Nets.
• Share weights to reduce number of parameters and
operations.
• Reduce total number of computations during model inference
by factoring the convolution layer
Optimization Pipeline
Copyright © 2017 Almond Research Pte Ltd. 3
• Feature Sharing
• “Distilling the Knowledge in a Neural Network”
• Very deep networks: Inception, Resnet, Highway networks, etc.
• Densely Connected Networks (2016)
• Resnet, Spatially Adaptive Computation Time (2017)
• Pruning, Weight Sharing, Compression
• “Deep Compression: Compressing Deep Neural Networks with
Pruning, Trained Quantization, And Huffman Coding
Relevant Research
Copyright © 2017 Almond Research Pte Ltd. 4
• 2D Convolutional Layer (multiply-add)
• Not counting activation, biases, batch norm
• Cost of Fully Convolutional Layers is a matrix multiply + bias
• Well studied: Pooling, Factorization
Review Computational Cost
[height*width]*(k/ stride)2
*(#inmaps*#outmaps)
Copyright © 2017 Almond Research Pte Ltd. 5
Model Error Rate
(Top 5)
Depth Parameters Runtime
(ms)
AlexNet1 19.9 8 62M 5.32
VGG2 9.33 16 144M 46.16
Inception2 9.15 22 6.8M 11.94
Resnet - 1012 6.05 110 1.7M 53.28
Size vs. Performance ImageNet
1 ImageNet 2012
2 Imagenet 2014
3 Pascal Titan X, Torch, CUDNN 5, Forward Pass Only
https://github.com/jcjohnson/cnn-benchmarks
Copyright © 2017 Almond Research Pte Ltd. 6
GoogleNet Inception
Reduce number of filters, before applying 3x3 and 5x5 convolutions
Copyright © 2017 Almond Research Pte Ltd. 7
Densely Connected Networks (Huang 2016)
• Factorize a convolutional layer into multiple (12) small blocks of filters.
• The input for each block is a concatenation of all previous convolutional filters.
• The output of each block is concatenated with the input.
• Recombine filters maps using a 1x1 convolution at the end of each stage.
Copyright © 2017 Almond Research Pte Ltd. 8
Densely Connected: Lowest Error On CIFAR 10/100
(millions)
Copyright © 2017 Almond Research Pte Ltd. 9
Example of a Real-time CNN
Copyright © 2017 Almond Research Pte Ltd. 10
• Kaggle Competition 2016 : Estimating Facial Keypoints
Real-time CNN : Finding Facial Keypoints
Training set of 7049 faces,
96x96 grayscale
Problem is to learn the (x,y)
position of 15 facial keypoints:
Eyes (6), Eyebrows (4), Nose, Lips (4)
Copyright © 2017 Almond Research Pte Ltd. 11
• Goal is to use all of the data, some incomplete, to train a bigger network
• Have to train a separate network for each subset of facial keypoints. Use
data augmentation (horizontal and vertical shifting).
• Solution is to share weights between the networks.
Dataset : Incomplete Data
Facial Keypoints Train (Shifting) Validation
Eyes (Corners) 1795 (6.6%) 449
Eyebrows 1753 (10%) 437
Mouth (Corners) 1801 (6.6%) 445
Smiley ☺ 5605 (10%) 1395
Complete 1717 (15%) 423
Copyright © 2017 Almond Research Pte Ltd. 12
• Train on 2100 faces, only the samples with all 15 features.
• Baseline: 6 layers (3 convolutional, 3 Fully Connected) and train on 1700
complete cases only.
• Network size is 32 – 64 – 128 – 500
Model I: Baseline Network
Copyright © 2017 Almond Research Pte Ltd. 13
Model II : Siamese Network
32 64 128 192
Concatenate output of
Last FC Layers of Each
Network
*Network is trained sequentially on each task
Copyright © 2017 Almond Research Pte Ltd. 14
Model III: Shared Densely Connected Network
Parameters: N=12, K=3
Each block learns 12 filters at a time. The output is concatenated to the input
to increase the depth of the network. N = 24 in the final stage of shared network.
Copyright © 2017 Almond Research Pte Ltd. 15
Convolution Input Output N K Image Size Ops Params
2x2 32 56 8 x3 94x94 92M 10,476
1x1 56 48 1/2 (L2 Weight
Regularization)
2x2 48 96 12 x4 47x47 82M 37,152
1x1 96 96 1/2
2x2 96 132 12 x3 23x23 28M 53,232
1x1 132 128
2x2 128 200 24 x3 23x23 70M 131,520
1x1 200 192 1/2
192 11x11
Densely Connected Model Parameters
Copyright © 2017 Almond Research Pte Ltd. 16
Model RMS Error
(Training)
RMS Error
(Testing)
Depth Operations Conv.
Params
Baseline 1.68 2.5 6 75M 61,760
Siamese 32-
64-128-192
1.80 2.07 9 192M 160,064
Siamese 48-
96-192-192
1.70 2.18 9 297M 286,176
Densely
Connected
2.13 2.29 23 366M 232,380
Results on Facial Keypoint Task
RMS = Root Mean Square
Copyright © 2017 Almond Research Pte Ltd. 17
• When a training set is small (1717 photos), must test whether the network
can actually perform the task in general. For feature tracking, measure
Translational Invariance: use a validation set with randomized
horizontal and vertical shifting.
• The Siamese model does the best with a Test RMS Error of 2.07. In
comparison, the baseline was the worst at generalization. Increasing the
amount of filters or adding more depth (Densely Connected) only
increased overfitting because there is not enough training data.
• The mean average error for Siamese model was already 0.75 pixels.
Targeting sub-pixel accuracy can also increase risk of overfitting or
“memorization”. Improving generalization is a more suitable goal.
Results Analysis
Copyright © 2017 Almond Research Pte Ltd. 18
Results (Siamese 32-64-128-192)
Predicted Ground Truth Predicted Ground Truth
Copyright © 2017 Almond Research Pte Ltd. 19
Frame Rates
Deployment Benchmarks
0.00
50.00
100.00
150.00
200.00
250.00
300.00
350.00
400.00
450.00
500.00
Baseline Siamese Dense
476.02
105.10
36.18
TX1 (Caffe)
TX1 (Caffe)
0.00
50.00
100.00
150.00
200.00
250.00
300.00
350.00
400.00
450.00
500.00
Baseline Siamese Dense
470.28
287.33
71.03
"Titan X (Theano)"
"Titan X (Theano)"
Batch size 32 Batch size 64
Copyright © 2017 Almond Research Pte Ltd. 20
• Ideally, we can train a very deep network with lots of training
data. In reality, we must be careful about overfitting.
• Siamese model is a way to increase the training set size
while sharing weights (convolutional filters) to reduce the
convolutional operations.
• For a Densely Connected model, depth can be increased
with a small increase in the total number of operations.
• Actual performance on the runtime platform may vary
depending on your choice of DNN library.
Conclusions
Copyright © 2017 Almond Research Pte Ltd. 21
• https://www.kaggle.com/c/facial-keypoints-detection
• http://danielnouri.org/notes/2014/12/17/using-convolutional-neural-nets-to-detect-
facial-keypoints-tutorial/
• Spatially Adaptive Computation Time for Residual Networks. Figurnov, Collins,
Zhu, Zhang, Huang, Vetrox, Salakhutdinov
• Deep Compression: Compressing Deep Neural Networks with Pruning, Trained
Quantization, And Huffman Coding
• Distilling the Knowledge in a Neural Network. Geoffrey Hinton, Oriol Vinyals, Jeff
Dean
• Going Deeper With Convolutions. Szegedy et Al.
• Densely Connected Convoutional Networks. Gao Huang, Zhuang Liu, Killian Q.
Weinberger
Resources
Copyright © 2017 Almond Research Pte Ltd. 22
• Conceptually complicated, but made of repeating blocks. At each
convolutional layer, only learn n * constant number of filters k.
[ (Input * k) + ((Input + k) * k) + … ((Input + (n-1)*k) * k) ]
• O(runtime) = (Input*(n*k) + (n2*k2) / 2) * (image size * kernel size)
• For fixed n*k < (input / 2),
• Convolutional Layer with output = input layers
• O(Input2 * (image size * kernel size))
Appendix : Analysis of Densely Connected
Networks
Copyright © 2017 Almond Research Pte Ltd. 23
• Reduce amount of memory required to store weights
(Pruning 10x, Weight Sharing 30x, Compression 50x)
• Also reduces Energy Consumption
• Speedup on Fully Connected Layers by Sparse Matrix
Multiply
Appendix: Weight Storage Compression

More Related Content

What's hot

“Improving Nursing Care with Privacy-Sensitive Edge Computer Vision,” a Prese...
“Improving Nursing Care with Privacy-Sensitive Edge Computer Vision,” a Prese...“Improving Nursing Care with Privacy-Sensitive Edge Computer Vision,” a Prese...
“Improving Nursing Care with Privacy-Sensitive Edge Computer Vision,” a Prese...
Edge AI and Vision Alliance
 
"Approaches for Energy Efficient Implementation of Deep Neural Networks," a P...
"Approaches for Energy Efficient Implementation of Deep Neural Networks," a P..."Approaches for Energy Efficient Implementation of Deep Neural Networks," a P...
"Approaches for Energy Efficient Implementation of Deep Neural Networks," a P...
Edge AI and Vision Alliance
 
“Introduction to DNN Model Compression Techniques,” a Presentation from Xailient
“Introduction to DNN Model Compression Techniques,” a Presentation from Xailient“Introduction to DNN Model Compression Techniques,” a Presentation from Xailient
“Introduction to DNN Model Compression Techniques,” a Presentation from Xailient
Edge AI and Vision Alliance
 
“Secure Hardware Architecture for Embedded Vision,” a Presentation from Neuro...
“Secure Hardware Architecture for Embedded Vision,” a Presentation from Neuro...“Secure Hardware Architecture for Embedded Vision,” a Presentation from Neuro...
“Secure Hardware Architecture for Embedded Vision,” a Presentation from Neuro...
Edge AI and Vision Alliance
 
"Deploying Deep Learning Models on Embedded Processors for Autonomous Systems...
"Deploying Deep Learning Models on Embedded Processors for Autonomous Systems..."Deploying Deep Learning Models on Embedded Processors for Autonomous Systems...
"Deploying Deep Learning Models on Embedded Processors for Autonomous Systems...
Edge AI and Vision Alliance
 
“Introduction to Simultaneous Localization and Mapping (SLAM),” a Presentatio...
“Introduction to Simultaneous Localization and Mapping (SLAM),” a Presentatio...“Introduction to Simultaneous Localization and Mapping (SLAM),” a Presentatio...
“Introduction to Simultaneous Localization and Mapping (SLAM),” a Presentatio...
Edge AI and Vision Alliance
 
Introducing google’s mobile nets
Introducing google’s mobile netsIntroducing google’s mobile nets
Introducing google’s mobile nets
Larry Guo
 
Machine Learning for Weather Forecasts
Machine Learning for Weather ForecastsMachine Learning for Weather Forecasts
Machine Learning for Weather Forecasts
inside-BigData.com
 
EfficientNet
EfficientNetEfficientNet
EfficientNet
Changjin Lee
 
"Machine Learning- based Image Compression: Ready for Prime Time?," a Present...
"Machine Learning- based Image Compression: Ready for Prime Time?," a Present..."Machine Learning- based Image Compression: Ready for Prime Time?," a Present...
"Machine Learning- based Image Compression: Ready for Prime Time?," a Present...
Edge AI and Vision Alliance
 
"Deep Learning for Manufacturing Inspection Applications," a Presentation fro...
"Deep Learning for Manufacturing Inspection Applications," a Presentation fro..."Deep Learning for Manufacturing Inspection Applications," a Presentation fro...
"Deep Learning for Manufacturing Inspection Applications," a Presentation fro...
Edge AI and Vision Alliance
 
“An Introduction to Single-Photon Avalanche Diodes—A New Type of Imager for C...
“An Introduction to Single-Photon Avalanche Diodes—A New Type of Imager for C...“An Introduction to Single-Photon Avalanche Diodes—A New Type of Imager for C...
“An Introduction to Single-Photon Avalanche Diodes—A New Type of Imager for C...
Edge AI and Vision Alliance
 
“Case Study: Facial Detection and Recognition for Always-On Applications,” a ...
“Case Study: Facial Detection and Recognition for Always-On Applications,” a ...“Case Study: Facial Detection and Recognition for Always-On Applications,” a ...
“Case Study: Facial Detection and Recognition for Always-On Applications,” a ...
Edge AI and Vision Alliance
 
"Quantizing Deep Networks for Efficient Inference at the Edge," a Presentatio...
"Quantizing Deep Networks for Efficient Inference at the Edge," a Presentatio..."Quantizing Deep Networks for Efficient Inference at the Edge," a Presentatio...
"Quantizing Deep Networks for Efficient Inference at the Edge," a Presentatio...
Edge AI and Vision Alliance
 
Enabling Power-Efficient AI Through Quantization
Enabling Power-Efficient AI Through QuantizationEnabling Power-Efficient AI Through Quantization
Enabling Power-Efficient AI Through Quantization
Qualcomm Research
 
“Efficient Deep Learning for 3D Point Cloud Understanding,” a Presentation fr...
“Efficient Deep Learning for 3D Point Cloud Understanding,” a Presentation fr...“Efficient Deep Learning for 3D Point Cloud Understanding,” a Presentation fr...
“Efficient Deep Learning for 3D Point Cloud Understanding,” a Presentation fr...
Edge AI and Vision Alliance
 
AI+ Remote Sensing: Applying Deep Learning to Image Enhancement, Analytics, a...
AI+ Remote Sensing: Applying Deep Learning to Image Enhancement, Analytics, a...AI+ Remote Sensing: Applying Deep Learning to Image Enhancement, Analytics, a...
AI+ Remote Sensing: Applying Deep Learning to Image Enhancement, Analytics, a...
Jui-Hsin (Larry) Lai
 
20211118 AI+ Remote Sensing
20211118 AI+ Remote Sensing20211118 AI+ Remote Sensing
20211118 AI+ Remote Sensing
Jui-Hsin (Larry) Lai
 
Joint super resolution and denoising from a single depth image
Joint super resolution and denoising from a single depth imageJoint super resolution and denoising from a single depth image
Joint super resolution and denoising from a single depth image
LogicMindtech Nologies
 
MobileNet Review | Mobile Net Research Paper Review | MobileNet v1 Paper Expl...
MobileNet Review | Mobile Net Research Paper Review | MobileNet v1 Paper Expl...MobileNet Review | Mobile Net Research Paper Review | MobileNet v1 Paper Expl...
MobileNet Review | Mobile Net Research Paper Review | MobileNet v1 Paper Expl...
Laxmi Kant Tiwari
 

What's hot (20)

“Improving Nursing Care with Privacy-Sensitive Edge Computer Vision,” a Prese...
“Improving Nursing Care with Privacy-Sensitive Edge Computer Vision,” a Prese...“Improving Nursing Care with Privacy-Sensitive Edge Computer Vision,” a Prese...
“Improving Nursing Care with Privacy-Sensitive Edge Computer Vision,” a Prese...
 
"Approaches for Energy Efficient Implementation of Deep Neural Networks," a P...
"Approaches for Energy Efficient Implementation of Deep Neural Networks," a P..."Approaches for Energy Efficient Implementation of Deep Neural Networks," a P...
"Approaches for Energy Efficient Implementation of Deep Neural Networks," a P...
 
“Introduction to DNN Model Compression Techniques,” a Presentation from Xailient
“Introduction to DNN Model Compression Techniques,” a Presentation from Xailient“Introduction to DNN Model Compression Techniques,” a Presentation from Xailient
“Introduction to DNN Model Compression Techniques,” a Presentation from Xailient
 
“Secure Hardware Architecture for Embedded Vision,” a Presentation from Neuro...
“Secure Hardware Architecture for Embedded Vision,” a Presentation from Neuro...“Secure Hardware Architecture for Embedded Vision,” a Presentation from Neuro...
“Secure Hardware Architecture for Embedded Vision,” a Presentation from Neuro...
 
"Deploying Deep Learning Models on Embedded Processors for Autonomous Systems...
"Deploying Deep Learning Models on Embedded Processors for Autonomous Systems..."Deploying Deep Learning Models on Embedded Processors for Autonomous Systems...
"Deploying Deep Learning Models on Embedded Processors for Autonomous Systems...
 
“Introduction to Simultaneous Localization and Mapping (SLAM),” a Presentatio...
“Introduction to Simultaneous Localization and Mapping (SLAM),” a Presentatio...“Introduction to Simultaneous Localization and Mapping (SLAM),” a Presentatio...
“Introduction to Simultaneous Localization and Mapping (SLAM),” a Presentatio...
 
Introducing google’s mobile nets
Introducing google’s mobile netsIntroducing google’s mobile nets
Introducing google’s mobile nets
 
Machine Learning for Weather Forecasts
Machine Learning for Weather ForecastsMachine Learning for Weather Forecasts
Machine Learning for Weather Forecasts
 
EfficientNet
EfficientNetEfficientNet
EfficientNet
 
"Machine Learning- based Image Compression: Ready for Prime Time?," a Present...
"Machine Learning- based Image Compression: Ready for Prime Time?," a Present..."Machine Learning- based Image Compression: Ready for Prime Time?," a Present...
"Machine Learning- based Image Compression: Ready for Prime Time?," a Present...
 
"Deep Learning for Manufacturing Inspection Applications," a Presentation fro...
"Deep Learning for Manufacturing Inspection Applications," a Presentation fro..."Deep Learning for Manufacturing Inspection Applications," a Presentation fro...
"Deep Learning for Manufacturing Inspection Applications," a Presentation fro...
 
“An Introduction to Single-Photon Avalanche Diodes—A New Type of Imager for C...
“An Introduction to Single-Photon Avalanche Diodes—A New Type of Imager for C...“An Introduction to Single-Photon Avalanche Diodes—A New Type of Imager for C...
“An Introduction to Single-Photon Avalanche Diodes—A New Type of Imager for C...
 
“Case Study: Facial Detection and Recognition for Always-On Applications,” a ...
“Case Study: Facial Detection and Recognition for Always-On Applications,” a ...“Case Study: Facial Detection and Recognition for Always-On Applications,” a ...
“Case Study: Facial Detection and Recognition for Always-On Applications,” a ...
 
"Quantizing Deep Networks for Efficient Inference at the Edge," a Presentatio...
"Quantizing Deep Networks for Efficient Inference at the Edge," a Presentatio..."Quantizing Deep Networks for Efficient Inference at the Edge," a Presentatio...
"Quantizing Deep Networks for Efficient Inference at the Edge," a Presentatio...
 
Enabling Power-Efficient AI Through Quantization
Enabling Power-Efficient AI Through QuantizationEnabling Power-Efficient AI Through Quantization
Enabling Power-Efficient AI Through Quantization
 
“Efficient Deep Learning for 3D Point Cloud Understanding,” a Presentation fr...
“Efficient Deep Learning for 3D Point Cloud Understanding,” a Presentation fr...“Efficient Deep Learning for 3D Point Cloud Understanding,” a Presentation fr...
“Efficient Deep Learning for 3D Point Cloud Understanding,” a Presentation fr...
 
AI+ Remote Sensing: Applying Deep Learning to Image Enhancement, Analytics, a...
AI+ Remote Sensing: Applying Deep Learning to Image Enhancement, Analytics, a...AI+ Remote Sensing: Applying Deep Learning to Image Enhancement, Analytics, a...
AI+ Remote Sensing: Applying Deep Learning to Image Enhancement, Analytics, a...
 
20211118 AI+ Remote Sensing
20211118 AI+ Remote Sensing20211118 AI+ Remote Sensing
20211118 AI+ Remote Sensing
 
Joint super resolution and denoising from a single depth image
Joint super resolution and denoising from a single depth imageJoint super resolution and denoising from a single depth image
Joint super resolution and denoising from a single depth image
 
MobileNet Review | Mobile Net Research Paper Review | MobileNet v1 Paper Expl...
MobileNet Review | Mobile Net Research Paper Review | MobileNet v1 Paper Expl...MobileNet Review | Mobile Net Research Paper Review | MobileNet v1 Paper Expl...
MobileNet Review | Mobile Net Research Paper Review | MobileNet v1 Paper Expl...
 

Similar to "Designing CNN Algorithms for Real-time Applications," a Presentation from Almond AI

“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...
“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...
“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...
Edge AI and Vision Alliance
 
Small Deep-Neural-Networks: Their Advantages and Their Design
Small Deep-Neural-Networks: Their Advantages and Their DesignSmall Deep-Neural-Networks: Their Advantages and Their Design
Small Deep-Neural-Networks: Their Advantages and Their Design
Forrest Iandola
 
PointNet
PointNetPointNet
Scalable Graph Convolutional Network Based Link Prediction on a Distributed G...
Scalable Graph Convolutional Network Based Link Prediction on a Distributed G...Scalable Graph Convolutional Network Based Link Prediction on a Distributed G...
Scalable Graph Convolutional Network Based Link Prediction on a Distributed G...
miyurud
 
AI On the Edge: Model Compression
AI On the Edge: Model CompressionAI On the Edge: Model Compression
AI On the Edge: Model Compression
Apache MXNet
 
Scaling Deep Learning Algorithms on Extreme Scale Architectures
Scaling Deep Learning Algorithms on Extreme Scale ArchitecturesScaling Deep Learning Algorithms on Extreme Scale Architectures
Scaling Deep Learning Algorithms on Extreme Scale Architectures
inside-BigData.com
 
Efficient de cvpr_2020_paper
Efficient de cvpr_2020_paperEfficient de cvpr_2020_paper
Efficient de cvpr_2020_paper
shanullah3
 
PR-183: MixNet: Mixed Depthwise Convolutional Kernels
PR-183: MixNet: Mixed Depthwise Convolutional KernelsPR-183: MixNet: Mixed Depthwise Convolutional Kernels
PR-183: MixNet: Mixed Depthwise Convolutional Kernels
Jinwon Lee
 
ML Paper Tutorial - Video Face Manipulation Detection Through Ensemble of CNN...
ML Paper Tutorial - Video Face Manipulation Detection Through Ensemble of CNN...ML Paper Tutorial - Video Face Manipulation Detection Through Ensemble of CNN...
ML Paper Tutorial - Video Face Manipulation Detection Through Ensemble of CNN...
Pei-Yuan Chien
 
Development of 3D convolutional neural network to recognize human activities ...
Development of 3D convolutional neural network to recognize human activities ...Development of 3D convolutional neural network to recognize human activities ...
Development of 3D convolutional neural network to recognize human activities ...
journalBEEI
 
AI optimizing HPC simulations (presentation from 6th EULAG Workshop)
AI optimizing HPC simulations (presentation from  6th EULAG Workshop)AI optimizing HPC simulations (presentation from  6th EULAG Workshop)
AI optimizing HPC simulations (presentation from 6th EULAG Workshop)
byteLAKE
 
ResNeSt: Split-Attention Networks
ResNeSt: Split-Attention NetworksResNeSt: Split-Attention Networks
ResNeSt: Split-Attention Networks
Seunghyun Hwang
 
RunPool: A Dynamic Pooling Layer for Convolution Neural Network
RunPool: A Dynamic Pooling Layer for Convolution Neural NetworkRunPool: A Dynamic Pooling Layer for Convolution Neural Network
RunPool: A Dynamic Pooling Layer for Convolution Neural Network
Putra Wanda
 
Accelerating stochastic gradient descent using adaptive mini batch size3
Accelerating stochastic gradient descent using adaptive mini batch size3Accelerating stochastic gradient descent using adaptive mini batch size3
Accelerating stochastic gradient descent using adaptive mini batch size3
muayyad alsadi
 
lec6a.ppt
lec6a.pptlec6a.ppt
lec6a.ppt
SaadMemon23
 
"Deep Learning Beyond Cats and Cars: Developing a Real-life DNN-based Embedde...
"Deep Learning Beyond Cats and Cars: Developing a Real-life DNN-based Embedde..."Deep Learning Beyond Cats and Cars: Developing a Real-life DNN-based Embedde...
"Deep Learning Beyond Cats and Cars: Developing a Real-life DNN-based Embedde...
Edge AI and Vision Alliance
 
Improving Hardware Efficiency for DNN Applications
Improving Hardware Efficiency for DNN ApplicationsImproving Hardware Efficiency for DNN Applications
Improving Hardware Efficiency for DNN Applications
Chester Chen
 
Computer Vision for Beginners
Computer Vision for BeginnersComputer Vision for Beginners
Computer Vision for Beginners
Sanghamitra Deb
 
IRJET - Factors Affecting Deployment of Deep Learning based Face Recognition ...
IRJET - Factors Affecting Deployment of Deep Learning based Face Recognition ...IRJET - Factors Affecting Deployment of Deep Learning based Face Recognition ...
IRJET - Factors Affecting Deployment of Deep Learning based Face Recognition ...
IRJET Journal
 
FAST ALGORITHMS FOR UNSUPERVISED LEARNING IN LARGE DATA SETS
FAST ALGORITHMS FOR UNSUPERVISED LEARNING IN LARGE DATA SETSFAST ALGORITHMS FOR UNSUPERVISED LEARNING IN LARGE DATA SETS
FAST ALGORITHMS FOR UNSUPERVISED LEARNING IN LARGE DATA SETS
csandit
 

Similar to "Designing CNN Algorithms for Real-time Applications," a Presentation from Almond AI (20)

“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...
“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...
“High-fidelity Conversion of Floating-point Networks for Low-precision Infere...
 
Small Deep-Neural-Networks: Their Advantages and Their Design
Small Deep-Neural-Networks: Their Advantages and Their DesignSmall Deep-Neural-Networks: Their Advantages and Their Design
Small Deep-Neural-Networks: Their Advantages and Their Design
 
PointNet
PointNetPointNet
PointNet
 
Scalable Graph Convolutional Network Based Link Prediction on a Distributed G...
Scalable Graph Convolutional Network Based Link Prediction on a Distributed G...Scalable Graph Convolutional Network Based Link Prediction on a Distributed G...
Scalable Graph Convolutional Network Based Link Prediction on a Distributed G...
 
AI On the Edge: Model Compression
AI On the Edge: Model CompressionAI On the Edge: Model Compression
AI On the Edge: Model Compression
 
Scaling Deep Learning Algorithms on Extreme Scale Architectures
Scaling Deep Learning Algorithms on Extreme Scale ArchitecturesScaling Deep Learning Algorithms on Extreme Scale Architectures
Scaling Deep Learning Algorithms on Extreme Scale Architectures
 
Efficient de cvpr_2020_paper
Efficient de cvpr_2020_paperEfficient de cvpr_2020_paper
Efficient de cvpr_2020_paper
 
PR-183: MixNet: Mixed Depthwise Convolutional Kernels
PR-183: MixNet: Mixed Depthwise Convolutional KernelsPR-183: MixNet: Mixed Depthwise Convolutional Kernels
PR-183: MixNet: Mixed Depthwise Convolutional Kernels
 
ML Paper Tutorial - Video Face Manipulation Detection Through Ensemble of CNN...
ML Paper Tutorial - Video Face Manipulation Detection Through Ensemble of CNN...ML Paper Tutorial - Video Face Manipulation Detection Through Ensemble of CNN...
ML Paper Tutorial - Video Face Manipulation Detection Through Ensemble of CNN...
 
Development of 3D convolutional neural network to recognize human activities ...
Development of 3D convolutional neural network to recognize human activities ...Development of 3D convolutional neural network to recognize human activities ...
Development of 3D convolutional neural network to recognize human activities ...
 
AI optimizing HPC simulations (presentation from 6th EULAG Workshop)
AI optimizing HPC simulations (presentation from  6th EULAG Workshop)AI optimizing HPC simulations (presentation from  6th EULAG Workshop)
AI optimizing HPC simulations (presentation from 6th EULAG Workshop)
 
ResNeSt: Split-Attention Networks
ResNeSt: Split-Attention NetworksResNeSt: Split-Attention Networks
ResNeSt: Split-Attention Networks
 
RunPool: A Dynamic Pooling Layer for Convolution Neural Network
RunPool: A Dynamic Pooling Layer for Convolution Neural NetworkRunPool: A Dynamic Pooling Layer for Convolution Neural Network
RunPool: A Dynamic Pooling Layer for Convolution Neural Network
 
Accelerating stochastic gradient descent using adaptive mini batch size3
Accelerating stochastic gradient descent using adaptive mini batch size3Accelerating stochastic gradient descent using adaptive mini batch size3
Accelerating stochastic gradient descent using adaptive mini batch size3
 
lec6a.ppt
lec6a.pptlec6a.ppt
lec6a.ppt
 
"Deep Learning Beyond Cats and Cars: Developing a Real-life DNN-based Embedde...
"Deep Learning Beyond Cats and Cars: Developing a Real-life DNN-based Embedde..."Deep Learning Beyond Cats and Cars: Developing a Real-life DNN-based Embedde...
"Deep Learning Beyond Cats and Cars: Developing a Real-life DNN-based Embedde...
 
Improving Hardware Efficiency for DNN Applications
Improving Hardware Efficiency for DNN ApplicationsImproving Hardware Efficiency for DNN Applications
Improving Hardware Efficiency for DNN Applications
 
Computer Vision for Beginners
Computer Vision for BeginnersComputer Vision for Beginners
Computer Vision for Beginners
 
IRJET - Factors Affecting Deployment of Deep Learning based Face Recognition ...
IRJET - Factors Affecting Deployment of Deep Learning based Face Recognition ...IRJET - Factors Affecting Deployment of Deep Learning based Face Recognition ...
IRJET - Factors Affecting Deployment of Deep Learning based Face Recognition ...
 
FAST ALGORITHMS FOR UNSUPERVISED LEARNING IN LARGE DATA SETS
FAST ALGORITHMS FOR UNSUPERVISED LEARNING IN LARGE DATA SETSFAST ALGORITHMS FOR UNSUPERVISED LEARNING IN LARGE DATA SETS
FAST ALGORITHMS FOR UNSUPERVISED LEARNING IN LARGE DATA SETS
 

More from Edge AI and Vision Alliance

“Temporal Event Neural Networks: A More Efficient Alternative to the Transfor...
“Temporal Event Neural Networks: A More Efficient Alternative to the Transfor...“Temporal Event Neural Networks: A More Efficient Alternative to the Transfor...
“Temporal Event Neural Networks: A More Efficient Alternative to the Transfor...
Edge AI and Vision Alliance
 
“Silicon Slip-Ups: The Ten Most Common Errors Processor Suppliers Make (Numbe...
“Silicon Slip-Ups: The Ten Most Common Errors Processor Suppliers Make (Numbe...“Silicon Slip-Ups: The Ten Most Common Errors Processor Suppliers Make (Numbe...
“Silicon Slip-Ups: The Ten Most Common Errors Processor Suppliers Make (Numbe...
Edge AI and Vision Alliance
 
“How Axelera AI Uses Digital Compute-in-memory to Deliver Fast and Energy-eff...
“How Axelera AI Uses Digital Compute-in-memory to Deliver Fast and Energy-eff...“How Axelera AI Uses Digital Compute-in-memory to Deliver Fast and Energy-eff...
“How Axelera AI Uses Digital Compute-in-memory to Deliver Fast and Energy-eff...
Edge AI and Vision Alliance
 
“How Arm’s Machine Learning Solution Enables Vision Transformers at the Edge,...
“How Arm’s Machine Learning Solution Enables Vision Transformers at the Edge,...“How Arm’s Machine Learning Solution Enables Vision Transformers at the Edge,...
“How Arm’s Machine Learning Solution Enables Vision Transformers at the Edge,...
Edge AI and Vision Alliance
 
“Nx EVOS: A New Enterprise Operating System for Video and Visual AI,” a Prese...
“Nx EVOS: A New Enterprise Operating System for Video and Visual AI,” a Prese...“Nx EVOS: A New Enterprise Operating System for Video and Visual AI,” a Prese...
“Nx EVOS: A New Enterprise Operating System for Video and Visual AI,” a Prese...
Edge AI and Vision Alliance
 
“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...
“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...
“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...
Edge AI and Vision Alliance
 
"OpenCV for High-performance, Low-power Vision Applications on Snapdragon," a...
"OpenCV for High-performance, Low-power Vision Applications on Snapdragon," a..."OpenCV for High-performance, Low-power Vision Applications on Snapdragon," a...
"OpenCV for High-performance, Low-power Vision Applications on Snapdragon," a...
Edge AI and Vision Alliance
 
“Deploying Large Models on the Edge: Success Stories and Challenges,” a Prese...
“Deploying Large Models on the Edge: Success Stories and Challenges,” a Prese...“Deploying Large Models on the Edge: Success Stories and Challenges,” a Prese...
“Deploying Large Models on the Edge: Success Stories and Challenges,” a Prese...
Edge AI and Vision Alliance
 
“Scaling Vision-based Edge AI Solutions: From Prototype to Global Deployment,...
“Scaling Vision-based Edge AI Solutions: From Prototype to Global Deployment,...“Scaling Vision-based Edge AI Solutions: From Prototype to Global Deployment,...
“Scaling Vision-based Edge AI Solutions: From Prototype to Global Deployment,...
Edge AI and Vision Alliance
 
“What’s Next in On-device Generative AI,” a Presentation from Qualcomm
“What’s Next in On-device Generative AI,” a Presentation from Qualcomm“What’s Next in On-device Generative AI,” a Presentation from Qualcomm
“What’s Next in On-device Generative AI,” a Presentation from Qualcomm
Edge AI and Vision Alliance
 
“Learning Compact DNN Models for Embedded Vision,” a Presentation from the Un...
“Learning Compact DNN Models for Embedded Vision,” a Presentation from the Un...“Learning Compact DNN Models for Embedded Vision,” a Presentation from the Un...
“Learning Compact DNN Models for Embedded Vision,” a Presentation from the Un...
Edge AI and Vision Alliance
 
“Introduction to Computer Vision with CNNs,” a Presentation from Mohammad Hag...
“Introduction to Computer Vision with CNNs,” a Presentation from Mohammad Hag...“Introduction to Computer Vision with CNNs,” a Presentation from Mohammad Hag...
“Introduction to Computer Vision with CNNs,” a Presentation from Mohammad Hag...
Edge AI and Vision Alliance
 
“Selecting Tools for Developing, Monitoring and Maintaining ML Models,” a Pre...
“Selecting Tools for Developing, Monitoring and Maintaining ML Models,” a Pre...“Selecting Tools for Developing, Monitoring and Maintaining ML Models,” a Pre...
“Selecting Tools for Developing, Monitoring and Maintaining ML Models,” a Pre...
Edge AI and Vision Alliance
 
“Building Accelerated GStreamer Applications for Video and Audio AI,” a Prese...
“Building Accelerated GStreamer Applications for Video and Audio AI,” a Prese...“Building Accelerated GStreamer Applications for Video and Audio AI,” a Prese...
“Building Accelerated GStreamer Applications for Video and Audio AI,” a Prese...
Edge AI and Vision Alliance
 
“Understanding, Selecting and Optimizing Object Detectors for Edge Applicatio...
“Understanding, Selecting and Optimizing Object Detectors for Edge Applicatio...“Understanding, Selecting and Optimizing Object Detectors for Edge Applicatio...
“Understanding, Selecting and Optimizing Object Detectors for Edge Applicatio...
Edge AI and Vision Alliance
 
“Introduction to Modern LiDAR for Machine Perception,” a Presentation from th...
“Introduction to Modern LiDAR for Machine Perception,” a Presentation from th...“Introduction to Modern LiDAR for Machine Perception,” a Presentation from th...
“Introduction to Modern LiDAR for Machine Perception,” a Presentation from th...
Edge AI and Vision Alliance
 
“Vision-language Representations for Robotics,” a Presentation from the Unive...
“Vision-language Representations for Robotics,” a Presentation from the Unive...“Vision-language Representations for Robotics,” a Presentation from the Unive...
“Vision-language Representations for Robotics,” a Presentation from the Unive...
Edge AI and Vision Alliance
 
“ADAS and AV Sensors: What’s Winning and Why?,” a Presentation from TechInsights
“ADAS and AV Sensors: What’s Winning and Why?,” a Presentation from TechInsights“ADAS and AV Sensors: What’s Winning and Why?,” a Presentation from TechInsights
“ADAS and AV Sensors: What’s Winning and Why?,” a Presentation from TechInsights
Edge AI and Vision Alliance
 
“Computer Vision in Sports: Scalable Solutions for Downmarkets,” a Presentati...
“Computer Vision in Sports: Scalable Solutions for Downmarkets,” a Presentati...“Computer Vision in Sports: Scalable Solutions for Downmarkets,” a Presentati...
“Computer Vision in Sports: Scalable Solutions for Downmarkets,” a Presentati...
Edge AI and Vision Alliance
 
“Detecting Data Drift in Image Classification Neural Networks,” a Presentatio...
“Detecting Data Drift in Image Classification Neural Networks,” a Presentatio...“Detecting Data Drift in Image Classification Neural Networks,” a Presentatio...
“Detecting Data Drift in Image Classification Neural Networks,” a Presentatio...
Edge AI and Vision Alliance
 

More from Edge AI and Vision Alliance (20)

“Temporal Event Neural Networks: A More Efficient Alternative to the Transfor...
“Temporal Event Neural Networks: A More Efficient Alternative to the Transfor...“Temporal Event Neural Networks: A More Efficient Alternative to the Transfor...
“Temporal Event Neural Networks: A More Efficient Alternative to the Transfor...
 
“Silicon Slip-Ups: The Ten Most Common Errors Processor Suppliers Make (Numbe...
“Silicon Slip-Ups: The Ten Most Common Errors Processor Suppliers Make (Numbe...“Silicon Slip-Ups: The Ten Most Common Errors Processor Suppliers Make (Numbe...
“Silicon Slip-Ups: The Ten Most Common Errors Processor Suppliers Make (Numbe...
 
“How Axelera AI Uses Digital Compute-in-memory to Deliver Fast and Energy-eff...
“How Axelera AI Uses Digital Compute-in-memory to Deliver Fast and Energy-eff...“How Axelera AI Uses Digital Compute-in-memory to Deliver Fast and Energy-eff...
“How Axelera AI Uses Digital Compute-in-memory to Deliver Fast and Energy-eff...
 
“How Arm’s Machine Learning Solution Enables Vision Transformers at the Edge,...
“How Arm’s Machine Learning Solution Enables Vision Transformers at the Edge,...“How Arm’s Machine Learning Solution Enables Vision Transformers at the Edge,...
“How Arm’s Machine Learning Solution Enables Vision Transformers at the Edge,...
 
“Nx EVOS: A New Enterprise Operating System for Video and Visual AI,” a Prese...
“Nx EVOS: A New Enterprise Operating System for Video and Visual AI,” a Prese...“Nx EVOS: A New Enterprise Operating System for Video and Visual AI,” a Prese...
“Nx EVOS: A New Enterprise Operating System for Video and Visual AI,” a Prese...
 
“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...
“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...
“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...
 
"OpenCV for High-performance, Low-power Vision Applications on Snapdragon," a...
"OpenCV for High-performance, Low-power Vision Applications on Snapdragon," a..."OpenCV for High-performance, Low-power Vision Applications on Snapdragon," a...
"OpenCV for High-performance, Low-power Vision Applications on Snapdragon," a...
 
“Deploying Large Models on the Edge: Success Stories and Challenges,” a Prese...
“Deploying Large Models on the Edge: Success Stories and Challenges,” a Prese...“Deploying Large Models on the Edge: Success Stories and Challenges,” a Prese...
“Deploying Large Models on the Edge: Success Stories and Challenges,” a Prese...
 
“Scaling Vision-based Edge AI Solutions: From Prototype to Global Deployment,...
“Scaling Vision-based Edge AI Solutions: From Prototype to Global Deployment,...“Scaling Vision-based Edge AI Solutions: From Prototype to Global Deployment,...
“Scaling Vision-based Edge AI Solutions: From Prototype to Global Deployment,...
 
“What’s Next in On-device Generative AI,” a Presentation from Qualcomm
“What’s Next in On-device Generative AI,” a Presentation from Qualcomm“What’s Next in On-device Generative AI,” a Presentation from Qualcomm
“What’s Next in On-device Generative AI,” a Presentation from Qualcomm
 
“Learning Compact DNN Models for Embedded Vision,” a Presentation from the Un...
“Learning Compact DNN Models for Embedded Vision,” a Presentation from the Un...“Learning Compact DNN Models for Embedded Vision,” a Presentation from the Un...
“Learning Compact DNN Models for Embedded Vision,” a Presentation from the Un...
 
“Introduction to Computer Vision with CNNs,” a Presentation from Mohammad Hag...
“Introduction to Computer Vision with CNNs,” a Presentation from Mohammad Hag...“Introduction to Computer Vision with CNNs,” a Presentation from Mohammad Hag...
“Introduction to Computer Vision with CNNs,” a Presentation from Mohammad Hag...
 
“Selecting Tools for Developing, Monitoring and Maintaining ML Models,” a Pre...
“Selecting Tools for Developing, Monitoring and Maintaining ML Models,” a Pre...“Selecting Tools for Developing, Monitoring and Maintaining ML Models,” a Pre...
“Selecting Tools for Developing, Monitoring and Maintaining ML Models,” a Pre...
 
“Building Accelerated GStreamer Applications for Video and Audio AI,” a Prese...
“Building Accelerated GStreamer Applications for Video and Audio AI,” a Prese...“Building Accelerated GStreamer Applications for Video and Audio AI,” a Prese...
“Building Accelerated GStreamer Applications for Video and Audio AI,” a Prese...
 
“Understanding, Selecting and Optimizing Object Detectors for Edge Applicatio...
“Understanding, Selecting and Optimizing Object Detectors for Edge Applicatio...“Understanding, Selecting and Optimizing Object Detectors for Edge Applicatio...
“Understanding, Selecting and Optimizing Object Detectors for Edge Applicatio...
 
“Introduction to Modern LiDAR for Machine Perception,” a Presentation from th...
“Introduction to Modern LiDAR for Machine Perception,” a Presentation from th...“Introduction to Modern LiDAR for Machine Perception,” a Presentation from th...
“Introduction to Modern LiDAR for Machine Perception,” a Presentation from th...
 
“Vision-language Representations for Robotics,” a Presentation from the Unive...
“Vision-language Representations for Robotics,” a Presentation from the Unive...“Vision-language Representations for Robotics,” a Presentation from the Unive...
“Vision-language Representations for Robotics,” a Presentation from the Unive...
 
“ADAS and AV Sensors: What’s Winning and Why?,” a Presentation from TechInsights
“ADAS and AV Sensors: What’s Winning and Why?,” a Presentation from TechInsights“ADAS and AV Sensors: What’s Winning and Why?,” a Presentation from TechInsights
“ADAS and AV Sensors: What’s Winning and Why?,” a Presentation from TechInsights
 
“Computer Vision in Sports: Scalable Solutions for Downmarkets,” a Presentati...
“Computer Vision in Sports: Scalable Solutions for Downmarkets,” a Presentati...“Computer Vision in Sports: Scalable Solutions for Downmarkets,” a Presentati...
“Computer Vision in Sports: Scalable Solutions for Downmarkets,” a Presentati...
 
“Detecting Data Drift in Image Classification Neural Networks,” a Presentatio...
“Detecting Data Drift in Image Classification Neural Networks,” a Presentatio...“Detecting Data Drift in Image Classification Neural Networks,” a Presentatio...
“Detecting Data Drift in Image Classification Neural Networks,” a Presentatio...
 

Recently uploaded

Recommendation System using RAG Architecture
Recommendation System using RAG ArchitectureRecommendation System using RAG Architecture
Recommendation System using RAG Architecture
fredae14
 
Energy Efficient Video Encoding for Cloud and Edge Computing Instances
Energy Efficient Video Encoding for Cloud and Edge Computing InstancesEnergy Efficient Video Encoding for Cloud and Edge Computing Instances
Energy Efficient Video Encoding for Cloud and Edge Computing Instances
Alpen-Adria-Universität
 
5th LF Energy Power Grid Model Meet-up Slides
5th LF Energy Power Grid Model Meet-up Slides5th LF Energy Power Grid Model Meet-up Slides
5th LF Energy Power Grid Model Meet-up Slides
DanBrown980551
 
UiPath Test Automation using UiPath Test Suite series, part 6
UiPath Test Automation using UiPath Test Suite series, part 6UiPath Test Automation using UiPath Test Suite series, part 6
UiPath Test Automation using UiPath Test Suite series, part 6
DianaGray10
 
Driving Business Innovation: Latest Generative AI Advancements & Success Story
Driving Business Innovation: Latest Generative AI Advancements & Success StoryDriving Business Innovation: Latest Generative AI Advancements & Success Story
Driving Business Innovation: Latest Generative AI Advancements & Success Story
Safe Software
 
みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...
みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...
みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...
名前 です男
 
Main news related to the CCS TSI 2023 (2023/1695)
Main news related to the CCS TSI 2023 (2023/1695)Main news related to the CCS TSI 2023 (2023/1695)
Main news related to the CCS TSI 2023 (2023/1695)
Jakub Marek
 
OpenID AuthZEN Interop Read Out - Authorization
OpenID AuthZEN Interop Read Out - AuthorizationOpenID AuthZEN Interop Read Out - Authorization
OpenID AuthZEN Interop Read Out - Authorization
David Brossard
 
Deep Dive: AI-Powered Marketing to Get More Leads and Customers with HyperGro...
Deep Dive: AI-Powered Marketing to Get More Leads and Customers with HyperGro...Deep Dive: AI-Powered Marketing to Get More Leads and Customers with HyperGro...
Deep Dive: AI-Powered Marketing to Get More Leads and Customers with HyperGro...
saastr
 
Introduction of Cybersecurity with OSS at Code Europe 2024
Introduction of Cybersecurity with OSS  at Code Europe 2024Introduction of Cybersecurity with OSS  at Code Europe 2024
Introduction of Cybersecurity with OSS at Code Europe 2024
Hiroshi SHIBATA
 
Presentation of the OECD Artificial Intelligence Review of Germany
Presentation of the OECD Artificial Intelligence Review of GermanyPresentation of the OECD Artificial Intelligence Review of Germany
Presentation of the OECD Artificial Intelligence Review of Germany
innovationoecd
 
How to Get CNIC Information System with Paksim Ga.pptx
How to Get CNIC Information System with Paksim Ga.pptxHow to Get CNIC Information System with Paksim Ga.pptx
How to Get CNIC Information System with Paksim Ga.pptx
danishmna97
 
How to use Firebase Data Connect For Flutter
How to use Firebase Data Connect For FlutterHow to use Firebase Data Connect For Flutter
How to use Firebase Data Connect For Flutter
Daiki Mogmet Ito
 
Your One-Stop Shop for Python Success: Top 10 US Python Development Providers
Your One-Stop Shop for Python Success: Top 10 US Python Development ProvidersYour One-Stop Shop for Python Success: Top 10 US Python Development Providers
Your One-Stop Shop for Python Success: Top 10 US Python Development Providers
akankshawande
 
Generating privacy-protected synthetic data using Secludy and Milvus
Generating privacy-protected synthetic data using Secludy and MilvusGenerating privacy-protected synthetic data using Secludy and Milvus
Generating privacy-protected synthetic data using Secludy and Milvus
Zilliz
 
Webinar: Designing a schema for a Data Warehouse
Webinar: Designing a schema for a Data WarehouseWebinar: Designing a schema for a Data Warehouse
Webinar: Designing a schema for a Data Warehouse
Federico Razzoli
 
Building Production Ready Search Pipelines with Spark and Milvus
Building Production Ready Search Pipelines with Spark and MilvusBuilding Production Ready Search Pipelines with Spark and Milvus
Building Production Ready Search Pipelines with Spark and Milvus
Zilliz
 
HCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAU
HCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAUHCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAU
HCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAU
panagenda
 
Taking AI to the Next Level in Manufacturing.pdf
Taking AI to the Next Level in Manufacturing.pdfTaking AI to the Next Level in Manufacturing.pdf
Taking AI to the Next Level in Manufacturing.pdf
ssuserfac0301
 
Let's Integrate MuleSoft RPA, COMPOSER, APM with AWS IDP along with Slack
Let's Integrate MuleSoft RPA, COMPOSER, APM with AWS IDP along with SlackLet's Integrate MuleSoft RPA, COMPOSER, APM with AWS IDP along with Slack
Let's Integrate MuleSoft RPA, COMPOSER, APM with AWS IDP along with Slack
shyamraj55
 

Recently uploaded (20)

Recommendation System using RAG Architecture
Recommendation System using RAG ArchitectureRecommendation System using RAG Architecture
Recommendation System using RAG Architecture
 
Energy Efficient Video Encoding for Cloud and Edge Computing Instances
Energy Efficient Video Encoding for Cloud and Edge Computing InstancesEnergy Efficient Video Encoding for Cloud and Edge Computing Instances
Energy Efficient Video Encoding for Cloud and Edge Computing Instances
 
5th LF Energy Power Grid Model Meet-up Slides
5th LF Energy Power Grid Model Meet-up Slides5th LF Energy Power Grid Model Meet-up Slides
5th LF Energy Power Grid Model Meet-up Slides
 
UiPath Test Automation using UiPath Test Suite series, part 6
UiPath Test Automation using UiPath Test Suite series, part 6UiPath Test Automation using UiPath Test Suite series, part 6
UiPath Test Automation using UiPath Test Suite series, part 6
 
Driving Business Innovation: Latest Generative AI Advancements & Success Story
Driving Business Innovation: Latest Generative AI Advancements & Success StoryDriving Business Innovation: Latest Generative AI Advancements & Success Story
Driving Business Innovation: Latest Generative AI Advancements & Success Story
 
みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...
みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...
みなさんこんにちはこれ何文字まで入るの?40文字以下不可とか本当に意味わからないけどこれ限界文字数書いてないからマジでやばい文字数いけるんじゃないの?えこ...
 
Main news related to the CCS TSI 2023 (2023/1695)
Main news related to the CCS TSI 2023 (2023/1695)Main news related to the CCS TSI 2023 (2023/1695)
Main news related to the CCS TSI 2023 (2023/1695)
 
OpenID AuthZEN Interop Read Out - Authorization
OpenID AuthZEN Interop Read Out - AuthorizationOpenID AuthZEN Interop Read Out - Authorization
OpenID AuthZEN Interop Read Out - Authorization
 
Deep Dive: AI-Powered Marketing to Get More Leads and Customers with HyperGro...
Deep Dive: AI-Powered Marketing to Get More Leads and Customers with HyperGro...Deep Dive: AI-Powered Marketing to Get More Leads and Customers with HyperGro...
Deep Dive: AI-Powered Marketing to Get More Leads and Customers with HyperGro...
 
Introduction of Cybersecurity with OSS at Code Europe 2024
Introduction of Cybersecurity with OSS  at Code Europe 2024Introduction of Cybersecurity with OSS  at Code Europe 2024
Introduction of Cybersecurity with OSS at Code Europe 2024
 
Presentation of the OECD Artificial Intelligence Review of Germany
Presentation of the OECD Artificial Intelligence Review of GermanyPresentation of the OECD Artificial Intelligence Review of Germany
Presentation of the OECD Artificial Intelligence Review of Germany
 
How to Get CNIC Information System with Paksim Ga.pptx
How to Get CNIC Information System with Paksim Ga.pptxHow to Get CNIC Information System with Paksim Ga.pptx
How to Get CNIC Information System with Paksim Ga.pptx
 
How to use Firebase Data Connect For Flutter
How to use Firebase Data Connect For FlutterHow to use Firebase Data Connect For Flutter
How to use Firebase Data Connect For Flutter
 
Your One-Stop Shop for Python Success: Top 10 US Python Development Providers
Your One-Stop Shop for Python Success: Top 10 US Python Development ProvidersYour One-Stop Shop for Python Success: Top 10 US Python Development Providers
Your One-Stop Shop for Python Success: Top 10 US Python Development Providers
 
Generating privacy-protected synthetic data using Secludy and Milvus
Generating privacy-protected synthetic data using Secludy and MilvusGenerating privacy-protected synthetic data using Secludy and Milvus
Generating privacy-protected synthetic data using Secludy and Milvus
 
Webinar: Designing a schema for a Data Warehouse
Webinar: Designing a schema for a Data WarehouseWebinar: Designing a schema for a Data Warehouse
Webinar: Designing a schema for a Data Warehouse
 
Building Production Ready Search Pipelines with Spark and Milvus
Building Production Ready Search Pipelines with Spark and MilvusBuilding Production Ready Search Pipelines with Spark and Milvus
Building Production Ready Search Pipelines with Spark and Milvus
 
HCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAU
HCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAUHCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAU
HCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAU
 
Taking AI to the Next Level in Manufacturing.pdf
Taking AI to the Next Level in Manufacturing.pdfTaking AI to the Next Level in Manufacturing.pdf
Taking AI to the Next Level in Manufacturing.pdf
 
Let's Integrate MuleSoft RPA, COMPOSER, APM with AWS IDP along with Slack
Let's Integrate MuleSoft RPA, COMPOSER, APM with AWS IDP along with SlackLet's Integrate MuleSoft RPA, COMPOSER, APM with AWS IDP along with Slack
Let's Integrate MuleSoft RPA, COMPOSER, APM with AWS IDP along with Slack
 

"Designing CNN Algorithms for Real-time Applications," a Presentation from Almond AI

  • 1. Copyright © 2017 Almond Research Pte Ltd. 1 Matthew Chiu May 2nd 2017 Designing CNN Algorithms For Real- time Applications
  • 2. Copyright © 2017 Almond Research Pte Ltd. 2 • Consider 3 stages: Dataset – Model – Deployment • Understand relationship between size (depth) and performance of Deep Neural Nets. • Share weights to reduce number of parameters and operations. • Reduce total number of computations during model inference by factoring the convolution layer Optimization Pipeline
  • 3. Copyright © 2017 Almond Research Pte Ltd. 3 • Feature Sharing • “Distilling the Knowledge in a Neural Network” • Very deep networks: Inception, Resnet, Highway networks, etc. • Densely Connected Networks (2016) • Resnet, Spatially Adaptive Computation Time (2017) • Pruning, Weight Sharing, Compression • “Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization, And Huffman Coding Relevant Research
  • 4. Copyright © 2017 Almond Research Pte Ltd. 4 • 2D Convolutional Layer (multiply-add) • Not counting activation, biases, batch norm • Cost of Fully Convolutional Layers is a matrix multiply + bias • Well studied: Pooling, Factorization Review Computational Cost [height*width]*(k/ stride)2 *(#inmaps*#outmaps)
  • 5. Copyright © 2017 Almond Research Pte Ltd. 5 Model Error Rate (Top 5) Depth Parameters Runtime (ms) AlexNet1 19.9 8 62M 5.32 VGG2 9.33 16 144M 46.16 Inception2 9.15 22 6.8M 11.94 Resnet - 1012 6.05 110 1.7M 53.28 Size vs. Performance ImageNet 1 ImageNet 2012 2 Imagenet 2014 3 Pascal Titan X, Torch, CUDNN 5, Forward Pass Only https://github.com/jcjohnson/cnn-benchmarks
  • 6. Copyright © 2017 Almond Research Pte Ltd. 6 GoogleNet Inception Reduce number of filters, before applying 3x3 and 5x5 convolutions
  • 7. Copyright © 2017 Almond Research Pte Ltd. 7 Densely Connected Networks (Huang 2016) • Factorize a convolutional layer into multiple (12) small blocks of filters. • The input for each block is a concatenation of all previous convolutional filters. • The output of each block is concatenated with the input. • Recombine filters maps using a 1x1 convolution at the end of each stage.
  • 8. Copyright © 2017 Almond Research Pte Ltd. 8 Densely Connected: Lowest Error On CIFAR 10/100 (millions)
  • 9. Copyright © 2017 Almond Research Pte Ltd. 9 Example of a Real-time CNN
  • 10. Copyright © 2017 Almond Research Pte Ltd. 10 • Kaggle Competition 2016 : Estimating Facial Keypoints Real-time CNN : Finding Facial Keypoints Training set of 7049 faces, 96x96 grayscale Problem is to learn the (x,y) position of 15 facial keypoints: Eyes (6), Eyebrows (4), Nose, Lips (4)
  • 11. Copyright © 2017 Almond Research Pte Ltd. 11 • Goal is to use all of the data, some incomplete, to train a bigger network • Have to train a separate network for each subset of facial keypoints. Use data augmentation (horizontal and vertical shifting). • Solution is to share weights between the networks. Dataset : Incomplete Data Facial Keypoints Train (Shifting) Validation Eyes (Corners) 1795 (6.6%) 449 Eyebrows 1753 (10%) 437 Mouth (Corners) 1801 (6.6%) 445 Smiley ☺ 5605 (10%) 1395 Complete 1717 (15%) 423
  • 12. Copyright © 2017 Almond Research Pte Ltd. 12 • Train on 2100 faces, only the samples with all 15 features. • Baseline: 6 layers (3 convolutional, 3 Fully Connected) and train on 1700 complete cases only. • Network size is 32 – 64 – 128 – 500 Model I: Baseline Network
  • 13. Copyright © 2017 Almond Research Pte Ltd. 13 Model II : Siamese Network 32 64 128 192 Concatenate output of Last FC Layers of Each Network *Network is trained sequentially on each task
  • 14. Copyright © 2017 Almond Research Pte Ltd. 14 Model III: Shared Densely Connected Network Parameters: N=12, K=3 Each block learns 12 filters at a time. The output is concatenated to the input to increase the depth of the network. N = 24 in the final stage of shared network.
  • 15. Copyright © 2017 Almond Research Pte Ltd. 15 Convolution Input Output N K Image Size Ops Params 2x2 32 56 8 x3 94x94 92M 10,476 1x1 56 48 1/2 (L2 Weight Regularization) 2x2 48 96 12 x4 47x47 82M 37,152 1x1 96 96 1/2 2x2 96 132 12 x3 23x23 28M 53,232 1x1 132 128 2x2 128 200 24 x3 23x23 70M 131,520 1x1 200 192 1/2 192 11x11 Densely Connected Model Parameters
  • 16. Copyright © 2017 Almond Research Pte Ltd. 16 Model RMS Error (Training) RMS Error (Testing) Depth Operations Conv. Params Baseline 1.68 2.5 6 75M 61,760 Siamese 32- 64-128-192 1.80 2.07 9 192M 160,064 Siamese 48- 96-192-192 1.70 2.18 9 297M 286,176 Densely Connected 2.13 2.29 23 366M 232,380 Results on Facial Keypoint Task RMS = Root Mean Square
  • 17. Copyright © 2017 Almond Research Pte Ltd. 17 • When a training set is small (1717 photos), must test whether the network can actually perform the task in general. For feature tracking, measure Translational Invariance: use a validation set with randomized horizontal and vertical shifting. • The Siamese model does the best with a Test RMS Error of 2.07. In comparison, the baseline was the worst at generalization. Increasing the amount of filters or adding more depth (Densely Connected) only increased overfitting because there is not enough training data. • The mean average error for Siamese model was already 0.75 pixels. Targeting sub-pixel accuracy can also increase risk of overfitting or “memorization”. Improving generalization is a more suitable goal. Results Analysis
  • 18. Copyright © 2017 Almond Research Pte Ltd. 18 Results (Siamese 32-64-128-192) Predicted Ground Truth Predicted Ground Truth
  • 19. Copyright © 2017 Almond Research Pte Ltd. 19 Frame Rates Deployment Benchmarks 0.00 50.00 100.00 150.00 200.00 250.00 300.00 350.00 400.00 450.00 500.00 Baseline Siamese Dense 476.02 105.10 36.18 TX1 (Caffe) TX1 (Caffe) 0.00 50.00 100.00 150.00 200.00 250.00 300.00 350.00 400.00 450.00 500.00 Baseline Siamese Dense 470.28 287.33 71.03 "Titan X (Theano)" "Titan X (Theano)" Batch size 32 Batch size 64
  • 20. Copyright © 2017 Almond Research Pte Ltd. 20 • Ideally, we can train a very deep network with lots of training data. In reality, we must be careful about overfitting. • Siamese model is a way to increase the training set size while sharing weights (convolutional filters) to reduce the convolutional operations. • For a Densely Connected model, depth can be increased with a small increase in the total number of operations. • Actual performance on the runtime platform may vary depending on your choice of DNN library. Conclusions
  • 21. Copyright © 2017 Almond Research Pte Ltd. 21 • https://www.kaggle.com/c/facial-keypoints-detection • http://danielnouri.org/notes/2014/12/17/using-convolutional-neural-nets-to-detect- facial-keypoints-tutorial/ • Spatially Adaptive Computation Time for Residual Networks. Figurnov, Collins, Zhu, Zhang, Huang, Vetrox, Salakhutdinov • Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization, And Huffman Coding • Distilling the Knowledge in a Neural Network. Geoffrey Hinton, Oriol Vinyals, Jeff Dean • Going Deeper With Convolutions. Szegedy et Al. • Densely Connected Convoutional Networks. Gao Huang, Zhuang Liu, Killian Q. Weinberger Resources
  • 22. Copyright © 2017 Almond Research Pte Ltd. 22 • Conceptually complicated, but made of repeating blocks. At each convolutional layer, only learn n * constant number of filters k. [ (Input * k) + ((Input + k) * k) + … ((Input + (n-1)*k) * k) ] • O(runtime) = (Input*(n*k) + (n2*k2) / 2) * (image size * kernel size) • For fixed n*k < (input / 2), • Convolutional Layer with output = input layers • O(Input2 * (image size * kernel size)) Appendix : Analysis of Densely Connected Networks
  • 23. Copyright © 2017 Almond Research Pte Ltd. 23 • Reduce amount of memory required to store weights (Pruning 10x, Weight Sharing 30x, Compression 50x) • Also reduces Energy Consumption • Speedup on Fully Connected Layers by Sparse Matrix Multiply Appendix: Weight Storage Compression