SlideShare a Scribd company logo
© 2019 xnor.ai
Methods for Creating Efficient
Convolutional Neural Networks
Mohammad Rastegari
xnor.ai
May 2019
© 2019 xnor.ai
Approaches to Efficient CNN
• Model design optimization
• Lower Precision (Quantization)
• Binary
• Sparse Models
• Lookup based CNN
• Compact Network Design
• Elastic model, Hierarchical convolution, Dimension-wise convolution
• Training Optimization
• Label Refinery
© 2019 xnor.ai
… …
Convolutional Neural Networks
© 2019 xnor.ai
• 1B – 20B FLOPs
Number of Operations :
• 0.25 – 3 fps
Inference time on CPU :
GPU !
*
+ − ×
© 2019 xnor.ai
Approaches to Efficient CNN
• Model design optimization
• Lower Precision (Quantization)
• Binary
• Sparse Models
• Lookup based CNN
• Compact Network Design
• Elastic model, Hierarchical convolution, Dimension-wise convolution
• Training Optimization
• Label Refinery
© 2019 xnor.ai
Lower Precision
32-bit 1-bit
Reducing Precision
• Saving Memory
• Saving Computation
{-1,+1} {0,1}
MUL XNOR
ADD, SUB Bit-Count (popcount)
8-bit
© 2019 xnor.ai
Why Binary?
Binary Instructions
• AND, OR, XOR, XNOR, PoPCount (Bit-Count)
Low Power Device
Easy to Implement in hardware
© 2019 xnor.ai
+ − × 1x 1x
Operations Memory Computation
+ − ~32x ~2x
XNOR
Bit-
count
~32x ~58x
Binary Weight Networks
XNOR-Networks
Theoretical Improvement
*
*
*
*
© 2019 xnor.ai
filter k 2 R , K = A ⇤k, where 8ij ki j = w ⇥h . K
or all sub-tensors in the input I . K i j corresponds to β for
e location ij (across width and height). This procedure is
ure2. Once weobtained thescaling factor ↵ for theweight
n I (denoted by K ), we can approximate the convolution
filter W mainly using binary operations:
W ⇡ (sign(I ) ~ sign(W )) K ↵ (11)
s the procedure explained in section 3.2 for approximating a convo-
ns.
onal vector where all of its enteries are 1. 1T
can be factored
and theoptimal solutions can beachieved from equation 2 as
sign(Y ) = sign(X T
) sign(W ) = H ⇤T
B ⇤ (9)
ependent, knowing that Y i = X i W i then,
= E [|X i |] E [|W i |] therefore,
P
|X i ||W i |
n
⇡
✓
1
n
kX k` 1
◆✓
1
n
kW k`1
◆
= β⇤
↵⇤
(10)
we convolve A with a 2D filter k 2 R , K = A ⇤k, where 8ij ki j = w ⇥h
contains scaling factors β for all sub-tensors in the input I . K i j corresponds to β
a sub-tensor centered at the location ij (across width and height). This procedur
shown in thethird row of figure2. Once weobtained thescaling factor ↵ for thewe
and β for all sub-tensors in I (denoted by K ), we can approximate the convolu
between input I and weight filter W mainly using binary operations:
I ⇤W ⇡ (sign(I ) ~ sign(W )) K ↵
lution using binary operations.
where 1 is an n-dimensional vector where all of its enteries are 1. 1T
ca
out from theoptimization and theoptimal solutions can beachieved from
follow
C⇤
= sign(Y ) = sign(X T
) sign(W ) = H ⇤T
B ⇤
Since |X i |, |W i | are independent, knowing that Y i = X i W i then,
E [|Y i |] = E [|X i ||W i |] = E [|X i |] E [|W i |] therefore,
γ⇤
=
P
|Y i |
n
=
P
|X i ||W i |
n
⇡
✓
1
n
kX k` 1
◆✓
1
n
kW k`1
◆
= β
c⇥w⇥h
WB
WB
WB = sign(W)
*
© 2019 xnor.ai
Quantization Error
WB = sign(W)
_ 0.75
WB
© 2019 xnor.ai
Optimal Scaling Factor
WB
© 2019 xnor.ai
Binary Input and Binary Weight (XNOR-Net)
a sub-tensor centered at the location ij (across width and height). This proced
shown in thethird row of figure2. Once weobtained thescaling factor ↵ for thew
and β for all sub-tensors in I (denoted by K ), we can approximate the convo
between input I and weight filter W mainly using binary operations:
I ⇤W ⇡ (sign(I ) ~ sign(W )) K ↵
i j
location ij (across width and height). This procedure is
re2. Once weobtained thescaling factor ↵ for theweight
I (denoted by K ), we can approximate the convolution
lter W mainly using binary operations:
⇡ (sign(I ) ~ sign(W )) K ↵ (11)
nal vector where all of its enteries are 1. 1T
can be factored
and theoptimal solutions can beachieved from equation 2 as
gn(Y ) = sign(X T
) sign(W ) = H ⇤T
B ⇤ (9)
endent, knowing that Y i = X i W i then,
E [|X i |] E [|W i |] therefore,
P
|X i ||W i |
n
⇡
✓
1
n
kX k` 1
◆✓
1
n
kW k`1
◆
= β⇤
↵⇤
(10)
volving weight filter W 2 Rc⇥w⇥h
(wherewi n w, hi n
WBXB
© 2019 xnor.ai
How to train a CNN with binary filters?
© 2019 xnor.ai
Training Binary Weight Networks
Naive Solution:
1. Train a network with real value parameters
2. Binarize the weight filters
© 2019 xnor.ai
0
10
20
30
40
50
60
70
80
Top-1 (%) ILSVRC2012
Full Precision
© 2019 xnor.ai
. . . . . .W
. . . . . .WB
Binarization
© 2019 xnor.ai
. . . . . .
Person
Dog
. . . . . .W
Binarization
© 2019 xnor.ai
Training XnorNet
W = W - ηGw
. . . . . .
. . . . . .
. . . . . .
Gw
W
Train for binary weights:
[XNOR-Networks, Rastegari et al, ECCV2016]
© 2019 xnor.ai
0
10
20
30
40
50
60
70
80
Top-1 (%) ILSVRC2012
0.2
[XNOR-Networks, Rastegari et al, ECCV2016]
© 2019 xnor.ai
Top-1 (%) ILSVRC2012
0.2
0
10
20
30
40
50
60
70
80
© 2019 xnor.ai
Approaches to Efficient CNN
• Model design optimization
• Lower Precision (Quantization)
• Binary (XNOR-Net)
• Sparse Models
• Lookup based CNN
• Compact Network Design
• Elastic model, Hierarchical convolution, Dimension-wise convolution
• Training Optimization
• Label Refinery
© 2019 xnor.ai
… …
Lookup Based CNN
© 2019 xnor.ai
© 2019 xnor.ai
© 2019 xnor.ai
*
© 2019 xnor.ai
How to train the discrete indexing?!!!!
*
© 2019 xnor.ai
© 2019 xnor.ai
*
© 2019 xnor.ai
0 10 20 30 40
Speed-up
AccuracyRate Image Classification
Few-shot Training
Few Iteration Training
OnDeviceTraining
© 2019 xnor.ai
Approaches to Efficient CNN
• Model design optimization
• Lower Precision (Quantization)
• Binary (XNOR-Net)
• Sparse Models
• Lookup based CNN
• Compact Network Design
• Elastic model, Hierarchical convolution, Dimension-wise convolution
• Training Optimization
• Label Refinery
© 2019 xnor.ai
Elastic: Instance Specific Efficiency
© 2019 xnor.ai
Challenging vs. Simple Images
© 2019 xnor.ai
Challenging vs. Simple Images
© 2019 xnor.ai
Challenging vs. Simple Images
© 2019 xnor.ai
Challenging vs. Simple Images
© 2019 xnor.ai
Challenging vs. Simple Images
© 2019 xnor.ai
ng Elas-
evaluate
eshow
ermore,
ransfer
classifi-
Xt [34],
be aug-
rch Im-
but no
tandard
Figure4: Imagenet Accuracy vs. FLOPSand ParametersThis
figure shows our Elastic model can achieve a lower error without
any extra (or with lower) computational cost.
parameters. Table2comparesthetop-1 andtop-5 error rates
of all of the base models with the Elastic augmentation (in-
© 2019 xnor.ai
Approaches to Efficient CNN
• Model design optimization
• Lower Precision (Quantization)
• Binary (XNOR-Net)
• Sparse Models
• Lookup based CNN
• Compact Network Design
• Elastic model, Hierarchical convolution, Dimension-wise convolution
• Training Optimization
• Label Refinery
© 2019 xnor.ai
Standard Convolution
Group Convolution
Depth-wise Convolution
Efficient Convolution
*
*
*
*
© 2019 xnor.ai
Standard Convolution
Dilated Convolution
© 2019 xnor.ai
Depth-wise Dilated Convolution (DDConv)
© 2019 xnor.ai
BNorm
Activ
Pool
Conv
BNorm
Activ
Pool
Conv
Standard CNN Block-structure
© 2019 xnor.ai
Mehtaet al.
(a)
RGB without HFF with HFF
(b)
Object Boundary Detection
Gridding Effect
Standard Block structure
© 2019 xnor.ai
Hierarchical DDConvs
© 2019 xnor.ai
Mehtaet al.
(a)
RGB without HFF with HFF
(b)Gridding Effect
Standard Block structure
No Gridding Effect
Hierarchical structure
Object Boundary Detection
© 2019 xnor.ai
Model FLOPs mIOU
HDDConv 1.4 B 69.1
DeepLabV3 2.84 B 71.8
Semantic Object Segmentation
© 2019 xnor.ai
Approaches to Efficient CNN
• Model design optimization
• Lower Precision (Quantization)
• Binary (XNOR-Net)
• Sparse Models
• Lookup based CNN
• Compact Network Design
• Elastic model, Hierarchical convolution, Dimension-wise convolution
• Training Optimization
• Label Refinery
© 2019 xnor.ai
Tensor Encoding Capacity
*
© 2019 xnor.ai
Tensor Encoding Capacity
*
© 2019 xnor.ai
Tensor Encoding Capacity
Ordering in channels has not being used for encoding
*
© 2019 xnor.ai
Tensor Encoding Capacity
Expensive Channel Fusion. O(c^2.h.w)Depth-Conv. O(c.h.w.k^2)
* *
© 2019 xnor.ai
Dimension-wise Convolution
Dim-Conv. O(c.h.w.k^2)
c
h
w
*
* *
© 2019 xnor.ai
Dimension-wise Convolution
Dim-Conv. O(c.h.w.k^2)
contains scaling factors β for all sub-tens
a sub-tensor centered at the location ij (
shown in thethird row of figure2. Oncew
and β for all sub-tensors in I (denoted b
between input I and weight filter W main
I ⇤W ⇡ (sign(I )
Efficient Channel Fusion. O(c^2+h.w)
contains scaling factors β for all sub-tensors in the inp
a sub-tensor centered at the location ij (across width
shown in thethird row of figure2. Onceweobtained the
and β for all sub-tensors in I (denoted by K ), we can
between input I and weight filter W mainly using binar
I ⇤W ⇡ (sign(I ) ~ sign(W ))
*
* *
© 2019 xnor.ai
© 2019 xnor.ai
FLOPs vs. Accuracy on Image Classification
ResNet-50
XNOR-res50
HDDconv
HDDConv
HDDConv
LCNN
LCNN
FLOPS(Log Millions)
ELASTIC
HDDConvDimConv
HDDConv
DimConv HDDConv
DimConv
Accuracy
© 2019 xnor.ai
Approaches to Efficient CNN
• Model design optimization
• Lower Precision (Quantization)
• Binary (XNOR-Net)
• Sparse Models
• Lookup based CNN
• Compact Network Design
• Elastic model, Hierarchical convolution, Dimension-wise convolution
• Training Optimization
• Label Refinery
© 2019 xnor.ai
Components in a Supervised Learning System
Data
• ImageNet, MSCOCO, SUN, …
• Data Augmentations
Model
• SVM, CNN
• Optimization Techniques (SGD,ADAM, RMSProp,…)
Label
• ?!!
© 2019 xnor.ai
Challenges with current labeling paradigm
Persian Cat
ball
• Incomplete
© 2019 xnor.ai
Challenges with current labeling paradigm
Random cropping
© 2019 xnor.ai
Challenges with current labeling paradigm
• Inconsistent
Dough
Butternut Squash
© 2019 xnor.ai
Challenges with current labeling paradigm
• Inconsistent
Dough
Butternut Squash
© 2019 xnor.ai
Challenges with current labeling paradigm
chrysanthemum dog
silky terrier
Car mirror
Sameamountofpenalization
• Taxonomy dependency
© 2019 xnor.ai
Labels should be:
Soft
Informative
Dynamic
Cat → 80%
Ball → 20%
Dog --> 60%
Cat --> 30%
Bear --> 10%
Dog --> 60%
Cat --> 10%
Bear --> 30%
Cat → 1 %
Ball → 99%
© 2019 xnor.ai
Label Refinery
Ground-
truth Label
Data
burrito burrito
plate
eggnog
burrito
plate
restaurant
Refinery
Top-1: 57.93 Top-1: 59.97 Top-1: 60.87 Top-1: 61.22
burrito
plate
restaurant
Refined Label
Data
Refinery
Refined Label
Data
Refinery
Refined Label
Data
Model
© 2019 xnor.ai
Label Refinery
Ground-
truth Label
Data
burrito burrito
plate
eggnog
burrito
plate
restaurant
Refinery
Top-1: 57.93 Top-1: 59.97 Top-1: 60.87 Top-1: 61.22
burrito
plate
restaurant
Refined Label
Data
Refinery
Refined Label
Data
Refinery
Refined Label
Data
Model
© 2019 xnor.ai
Label Refinery
Ground-
truth Label
Data
burrito burrito
plate
eggnog
burrito
plate
restaurant
Refinery
Top-1: 57.93 Top-1: 59.97 Top-1: 60.87 Top-1: 61.22
burrito
plate
restaurant
Refined Label
Data
Refinery
Refined Label
Data
Refinery
Refined Label
Data
Model
© 2019 xnor.ai
Model Top-1 Top-5 Top-1 Top-5
AlexNet 57.93 79.41 66.28 86.13
MobileNet-1 68.53 88.14 73.39 91.07
MobileNet-0.75 65.93 86.28 70.92 89.68
MobileNet-0.5 63.03 84.55 66.66 87.07
MobileNet-0.25 50.65 74.42 54.62 77.92
ResNet-50 75.7 92.81 76.5 93.12
ResNet-34 73.39 91.32 75.06 92.35
ResNet-18 69.7 89.26 72.52 90.73
ResNetXnor-50 63.1 83.61 73.31 89.18
VGG16 70.1 88.54 75 92.22
VGG19 71.39 89.44 75.46 92.52
DarkNet19 70.6 89.13 74.47 91.94
Label RefineryStandard Training
© 2019 xnor.ai
Model Top-1 Top-5 Top-1 Top-5
AlexNet 57.93 79.41 66.28 86.13
MobileNet-1 68.53 88.14 73.39 91.07
MobileNet-0.75 65.93 86.28 70.92 89.68
MobileNet-0.5 63.03 84.55 66.66 87.07
MobileNet-0.25 50.65 74.42 54.62 77.92
ResNet-50 75.7 92.81 76.5 93.12
ResNet-34 73.39 91.32 75.06 92.35
ResNet-18 69.7 89.26 72.52 90.73
ResNetXnor-50 63.1 83.61 73.31 89.18
VGG16 70.1 88.54 75 92.22
VGG19 71.39 89.44 75.46 92.52
DarkNet19 70.6 89.13 74.47 91.94
Label RefineryStandard Training
© 2019 xnor.ai 74
How far we can get with this efficiency?
Server & CloudEdge & Embedded Devices Mobile Devices
Compute Capability & Price HighLow
Traditional Home of AI
FPGA AI/Neural Accelerator GPU
CPUs
© 2019 xnor.ai
Thank You !!!
1. ELASTIC: Improving CNNs with Instance Specific Scaling PoliciesH Wang, A Kembhavi, A Farhadi, A Yuille, M Rastegari arXiv
preprint arXiv:1812.05262
2. ESPNetv2: A Light-weight, Power Efficient, and General Purpose Convolutional Neural NetworkS Mehta, M Rastegari, L
Shapiro, H Hajishirzi arXiv preprint arXiv:1811.11431
3. Label refinery: Improving imagenet classification through label progressionH Bagherinezhad, M Horton, M Rastegari, A
Farhadi arXiv preprint arXiv:1805.02641
4. Espnet: Efficient spatial pyramid of dilated convolutions for semantic segmentationS Mehta, M Rastegari, A Caspi, L
Shapiro, H Hajishirzi . Proceedings of the European Conference on Computer Vision (ECCV), 552-568
5. Xnor-net: Imagenet classification using binary convolutional neural networksM Rastegari, V Ordonez, J Redmon, A Farhadi
European Conference on Computer Vision, 525-542
6. Lcnn: Lookup-based convolutional neural networkH Bagherinezhad, M Rastegari, A Farhadi. Proceedings of the IEEE
Conference on Computer Vision and Pattern …

More Related Content

What's hot

Modification on Energy Efficient Design of DVB-T2 Constellation De-mapper
Modification on Energy Efficient Design of DVB-T2 Constellation De-mapperModification on Energy Efficient Design of DVB-T2 Constellation De-mapper
Modification on Energy Efficient Design of DVB-T2 Constellation De-mapper
IJERA Editor
 
D0532025
D0532025D0532025
D0532025
IOSR Journals
 
COSA and CSA based 32 -bit unsigned multipler
COSA and CSA based 32 -bit unsigned multiplerCOSA and CSA based 32 -bit unsigned multipler
COSA and CSA based 32 -bit unsigned multipler
inventy
 
Fixed point scaling
Fixed point scalingFixed point scaling
Fixed point scaling
rishi ram khanal
 
Impact of facts devices on zonal congestion management in deregulated power s...
Impact of facts devices on zonal congestion management in deregulated power s...Impact of facts devices on zonal congestion management in deregulated power s...
Impact of facts devices on zonal congestion management in deregulated power s...
Alexander Decker
 
Welcome to International Journal of Engineering Research and Development (IJERD)
Welcome to International Journal of Engineering Research and Development (IJERD)Welcome to International Journal of Engineering Research and Development (IJERD)
Welcome to International Journal of Engineering Research and Development (IJERD)
IJERD Editor
 
Graphical Model Selection for Big Data
Graphical Model Selection for Big DataGraphical Model Selection for Big Data
Graphical Model Selection for Big Data
Alexander Jung
 
Project seminar ppt_steelcasting
Project seminar ppt_steelcastingProject seminar ppt_steelcasting
Project seminar ppt_steelcasting
Rudra Narayan Paul
 
Parallel algorithm for computing edt with new architecture
Parallel algorithm for computing edt with new architectureParallel algorithm for computing edt with new architecture
Parallel algorithm for computing edt with new architecture
IAEME Publication
 
Digital image processing using matlab: basic transformations, filters and ope...
Digital image processing using matlab: basic transformations, filters and ope...Digital image processing using matlab: basic transformations, filters and ope...
Digital image processing using matlab: basic transformations, filters and ope...
thanh nguyen
 
Coursera: Structure Standing Still: Project Presentation
Coursera: Structure Standing Still: Project Presentation Coursera: Structure Standing Still: Project Presentation
Coursera: Structure Standing Still: Project Presentation
Creative Venture
 
A Comparative Study of Histogram Equalization Based Image Enhancement Techniq...
A Comparative Study of Histogram Equalization Based Image Enhancement Techniq...A Comparative Study of Histogram Equalization Based Image Enhancement Techniq...
A Comparative Study of Histogram Equalization Based Image Enhancement Techniq...
Shahbaz Alam
 
11.impact of facts devices on zonal congestion management in deregulated powe...
11.impact of facts devices on zonal congestion management in deregulated powe...11.impact of facts devices on zonal congestion management in deregulated powe...
11.impact of facts devices on zonal congestion management in deregulated powe...
Alexander Decker
 
Image restoration and reconstruction
Image restoration and reconstructionImage restoration and reconstruction
Image processing fundamentals
Image processing fundamentalsImage processing fundamentals
Image processing fundamentals
chandra mohan
 
Colorization with total variantion regularization
Colorization with total variantion regularizationColorization with total variantion regularization
Colorization with total variantion regularization
JÚLIO PEIXOTO
 
Log Transformation in Image Processing with Example
Log Transformation in Image Processing with ExampleLog Transformation in Image Processing with Example
Log Transformation in Image Processing with Example
Mustak Ahmmed
 
05 histogram processing DIP
05 histogram processing DIP05 histogram processing DIP
05 histogram processing DIP
babak danyal
 
carry maskable look ahead adder for approximate computing
carry maskable look ahead adder for approximate computingcarry maskable look ahead adder for approximate computing
carry maskable look ahead adder for approximate computing
Mahesh Dhava
 

What's hot (19)

Modification on Energy Efficient Design of DVB-T2 Constellation De-mapper
Modification on Energy Efficient Design of DVB-T2 Constellation De-mapperModification on Energy Efficient Design of DVB-T2 Constellation De-mapper
Modification on Energy Efficient Design of DVB-T2 Constellation De-mapper
 
D0532025
D0532025D0532025
D0532025
 
COSA and CSA based 32 -bit unsigned multipler
COSA and CSA based 32 -bit unsigned multiplerCOSA and CSA based 32 -bit unsigned multipler
COSA and CSA based 32 -bit unsigned multipler
 
Fixed point scaling
Fixed point scalingFixed point scaling
Fixed point scaling
 
Impact of facts devices on zonal congestion management in deregulated power s...
Impact of facts devices on zonal congestion management in deregulated power s...Impact of facts devices on zonal congestion management in deregulated power s...
Impact of facts devices on zonal congestion management in deregulated power s...
 
Welcome to International Journal of Engineering Research and Development (IJERD)
Welcome to International Journal of Engineering Research and Development (IJERD)Welcome to International Journal of Engineering Research and Development (IJERD)
Welcome to International Journal of Engineering Research and Development (IJERD)
 
Graphical Model Selection for Big Data
Graphical Model Selection for Big DataGraphical Model Selection for Big Data
Graphical Model Selection for Big Data
 
Project seminar ppt_steelcasting
Project seminar ppt_steelcastingProject seminar ppt_steelcasting
Project seminar ppt_steelcasting
 
Parallel algorithm for computing edt with new architecture
Parallel algorithm for computing edt with new architectureParallel algorithm for computing edt with new architecture
Parallel algorithm for computing edt with new architecture
 
Digital image processing using matlab: basic transformations, filters and ope...
Digital image processing using matlab: basic transformations, filters and ope...Digital image processing using matlab: basic transformations, filters and ope...
Digital image processing using matlab: basic transformations, filters and ope...
 
Coursera: Structure Standing Still: Project Presentation
Coursera: Structure Standing Still: Project Presentation Coursera: Structure Standing Still: Project Presentation
Coursera: Structure Standing Still: Project Presentation
 
A Comparative Study of Histogram Equalization Based Image Enhancement Techniq...
A Comparative Study of Histogram Equalization Based Image Enhancement Techniq...A Comparative Study of Histogram Equalization Based Image Enhancement Techniq...
A Comparative Study of Histogram Equalization Based Image Enhancement Techniq...
 
11.impact of facts devices on zonal congestion management in deregulated powe...
11.impact of facts devices on zonal congestion management in deregulated powe...11.impact of facts devices on zonal congestion management in deregulated powe...
11.impact of facts devices on zonal congestion management in deregulated powe...
 
Image restoration and reconstruction
Image restoration and reconstructionImage restoration and reconstruction
Image restoration and reconstruction
 
Image processing fundamentals
Image processing fundamentalsImage processing fundamentals
Image processing fundamentals
 
Colorization with total variantion regularization
Colorization with total variantion regularizationColorization with total variantion regularization
Colorization with total variantion regularization
 
Log Transformation in Image Processing with Example
Log Transformation in Image Processing with ExampleLog Transformation in Image Processing with Example
Log Transformation in Image Processing with Example
 
05 histogram processing DIP
05 histogram processing DIP05 histogram processing DIP
05 histogram processing DIP
 
carry maskable look ahead adder for approximate computing
carry maskable look ahead adder for approximate computingcarry maskable look ahead adder for approximate computing
carry maskable look ahead adder for approximate computing
 

Similar to "Methods for Creating Efficient Convolutional Neural Networks," a Presentation from Xnor.ai

International Journal of Engineering Research and Development (IJERD)
International Journal of Engineering Research and Development (IJERD)International Journal of Engineering Research and Development (IJERD)
International Journal of Engineering Research and Development (IJERD)
IJERD Editor
 
Distortion Correction Scheme for Multiresolution Camera Images
Distortion Correction Scheme for Multiresolution Camera ImagesDistortion Correction Scheme for Multiresolution Camera Images
Distortion Correction Scheme for Multiresolution Camera Images
Associate Professor in VSB Coimbatore
 
Design of 8-Bit Comparator Using 45nm CMOS Technology
Design of 8-Bit Comparator Using 45nm CMOS TechnologyDesign of 8-Bit Comparator Using 45nm CMOS Technology
Design of 8-Bit Comparator Using 45nm CMOS Technology
IJMER
 
Ax03303120316
Ax03303120316Ax03303120316
Ax03303120316
ijceronline
 
A Configurable and Low Power Hard-Decision Viterbi Decoder in VLSI Architecture
A Configurable and Low Power Hard-Decision Viterbi Decoder in VLSI ArchitectureA Configurable and Low Power Hard-Decision Viterbi Decoder in VLSI Architecture
A Configurable and Low Power Hard-Decision Viterbi Decoder in VLSI Architecture
IRJET Journal
 
Paper id 37201520
Paper id 37201520Paper id 37201520
Paper id 37201520
IJRAT
 
Computer Graphics
Computer GraphicsComputer Graphics
Computer Graphics
Griffinder VinHai
 
HARDWARE EFFICIENT SCALING FREE VECTORING AND ROTATIONAL CORDIC FOR DSP APPLI...
HARDWARE EFFICIENT SCALING FREE VECTORING AND ROTATIONAL CORDIC FOR DSP APPLI...HARDWARE EFFICIENT SCALING FREE VECTORING AND ROTATIONAL CORDIC FOR DSP APPLI...
HARDWARE EFFICIENT SCALING FREE VECTORING AND ROTATIONAL CORDIC FOR DSP APPLI...
VLSICS Design
 
A guide to Tensor and its applications in Machine Learning.pdf
A guide to Tensor and its applications in Machine Learning.pdfA guide to Tensor and its applications in Machine Learning.pdf
A guide to Tensor and its applications in Machine Learning.pdf
Vanessa Bridge
 
IOEfficientParalleMatrixMultiplication_present
IOEfficientParalleMatrixMultiplication_presentIOEfficientParalleMatrixMultiplication_present
IOEfficientParalleMatrixMultiplication_present
Shubham Joshi
 
I3602061067
I3602061067I3602061067
I3602061067
ijceronline
 
6 10
6 106 10
W4408123126
W4408123126W4408123126
W4408123126
IJERA Editor
 
Robust watermarking technique sppt
Robust watermarking technique spptRobust watermarking technique sppt
Robust watermarking technique sppt
Vijayakumar Veeramuthu
 
Design and Implementation of Different types of Carry skip adder
Design and Implementation of Different types of Carry skip adderDesign and Implementation of Different types of Carry skip adder
Design and Implementation of Different types of Carry skip adder
IRJET Journal
 
Lecture4 xing
Lecture4 xingLecture4 xing
Lecture4 xing
Tianlu Wang
 
Vector-Based Back Propagation Algorithm of.pdf
Vector-Based Back Propagation Algorithm of.pdfVector-Based Back Propagation Algorithm of.pdf
Vector-Based Back Propagation Algorithm of.pdf
Nesrine Wagaa
 
Performance Comparison of K-means Codebook Optimization using different Clust...
Performance Comparison of K-means Codebook Optimization using different Clust...Performance Comparison of K-means Codebook Optimization using different Clust...
Performance Comparison of K-means Codebook Optimization using different Clust...
IOSR Journals
 
I43024751
I43024751I43024751
I43024751
IJERA Editor
 
P410498102
P410498102P410498102
P410498102
IJERA Editor
 

Similar to "Methods for Creating Efficient Convolutional Neural Networks," a Presentation from Xnor.ai (20)

International Journal of Engineering Research and Development (IJERD)
International Journal of Engineering Research and Development (IJERD)International Journal of Engineering Research and Development (IJERD)
International Journal of Engineering Research and Development (IJERD)
 
Distortion Correction Scheme for Multiresolution Camera Images
Distortion Correction Scheme for Multiresolution Camera ImagesDistortion Correction Scheme for Multiresolution Camera Images
Distortion Correction Scheme for Multiresolution Camera Images
 
Design of 8-Bit Comparator Using 45nm CMOS Technology
Design of 8-Bit Comparator Using 45nm CMOS TechnologyDesign of 8-Bit Comparator Using 45nm CMOS Technology
Design of 8-Bit Comparator Using 45nm CMOS Technology
 
Ax03303120316
Ax03303120316Ax03303120316
Ax03303120316
 
A Configurable and Low Power Hard-Decision Viterbi Decoder in VLSI Architecture
A Configurable and Low Power Hard-Decision Viterbi Decoder in VLSI ArchitectureA Configurable and Low Power Hard-Decision Viterbi Decoder in VLSI Architecture
A Configurable and Low Power Hard-Decision Viterbi Decoder in VLSI Architecture
 
Paper id 37201520
Paper id 37201520Paper id 37201520
Paper id 37201520
 
Computer Graphics
Computer GraphicsComputer Graphics
Computer Graphics
 
HARDWARE EFFICIENT SCALING FREE VECTORING AND ROTATIONAL CORDIC FOR DSP APPLI...
HARDWARE EFFICIENT SCALING FREE VECTORING AND ROTATIONAL CORDIC FOR DSP APPLI...HARDWARE EFFICIENT SCALING FREE VECTORING AND ROTATIONAL CORDIC FOR DSP APPLI...
HARDWARE EFFICIENT SCALING FREE VECTORING AND ROTATIONAL CORDIC FOR DSP APPLI...
 
A guide to Tensor and its applications in Machine Learning.pdf
A guide to Tensor and its applications in Machine Learning.pdfA guide to Tensor and its applications in Machine Learning.pdf
A guide to Tensor and its applications in Machine Learning.pdf
 
IOEfficientParalleMatrixMultiplication_present
IOEfficientParalleMatrixMultiplication_presentIOEfficientParalleMatrixMultiplication_present
IOEfficientParalleMatrixMultiplication_present
 
I3602061067
I3602061067I3602061067
I3602061067
 
6 10
6 106 10
6 10
 
W4408123126
W4408123126W4408123126
W4408123126
 
Robust watermarking technique sppt
Robust watermarking technique spptRobust watermarking technique sppt
Robust watermarking technique sppt
 
Design and Implementation of Different types of Carry skip adder
Design and Implementation of Different types of Carry skip adderDesign and Implementation of Different types of Carry skip adder
Design and Implementation of Different types of Carry skip adder
 
Lecture4 xing
Lecture4 xingLecture4 xing
Lecture4 xing
 
Vector-Based Back Propagation Algorithm of.pdf
Vector-Based Back Propagation Algorithm of.pdfVector-Based Back Propagation Algorithm of.pdf
Vector-Based Back Propagation Algorithm of.pdf
 
Performance Comparison of K-means Codebook Optimization using different Clust...
Performance Comparison of K-means Codebook Optimization using different Clust...Performance Comparison of K-means Codebook Optimization using different Clust...
Performance Comparison of K-means Codebook Optimization using different Clust...
 
I43024751
I43024751I43024751
I43024751
 
P410498102
P410498102P410498102
P410498102
 

More from Edge AI and Vision Alliance

“Optimized Vision Language Models for Intelligent Transportation System Appli...
“Optimized Vision Language Models for Intelligent Transportation System Appli...“Optimized Vision Language Models for Intelligent Transportation System Appli...
“Optimized Vision Language Models for Intelligent Transportation System Appli...
Edge AI and Vision Alliance
 
“Image Signal Processing Optimization for Object Detection,” a Presentation f...
“Image Signal Processing Optimization for Object Detection,” a Presentation f...“Image Signal Processing Optimization for Object Detection,” a Presentation f...
“Image Signal Processing Optimization for Object Detection,” a Presentation f...
Edge AI and Vision Alliance
 
“Squeezing the Last Milliwatt and Cubic Millimeter from Smart Cameras Using t...
“Squeezing the Last Milliwatt and Cubic Millimeter from Smart Cameras Using t...“Squeezing the Last Milliwatt and Cubic Millimeter from Smart Cameras Using t...
“Squeezing the Last Milliwatt and Cubic Millimeter from Smart Cameras Using t...
Edge AI and Vision Alliance
 
"Maximize Your AI Compatibility with Flexible Pre- and Post-processing," a Pr...
"Maximize Your AI Compatibility with Flexible Pre- and Post-processing," a Pr..."Maximize Your AI Compatibility with Flexible Pre- and Post-processing," a Pr...
"Maximize Your AI Compatibility with Flexible Pre- and Post-processing," a Pr...
Edge AI and Vision Alliance
 
“Addressing Tomorrow’s Sensor Fusion and Processing Needs with Cadence’s Newe...
“Addressing Tomorrow’s Sensor Fusion and Processing Needs with Cadence’s Newe...“Addressing Tomorrow’s Sensor Fusion and Processing Needs with Cadence’s Newe...
“Addressing Tomorrow’s Sensor Fusion and Processing Needs with Cadence’s Newe...
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
 

More from Edge AI and Vision Alliance (20)

“Optimized Vision Language Models for Intelligent Transportation System Appli...
“Optimized Vision Language Models for Intelligent Transportation System Appli...“Optimized Vision Language Models for Intelligent Transportation System Appli...
“Optimized Vision Language Models for Intelligent Transportation System Appli...
 
“Image Signal Processing Optimization for Object Detection,” a Presentation f...
“Image Signal Processing Optimization for Object Detection,” a Presentation f...“Image Signal Processing Optimization for Object Detection,” a Presentation f...
“Image Signal Processing Optimization for Object Detection,” a Presentation f...
 
“Squeezing the Last Milliwatt and Cubic Millimeter from Smart Cameras Using t...
“Squeezing the Last Milliwatt and Cubic Millimeter from Smart Cameras Using t...“Squeezing the Last Milliwatt and Cubic Millimeter from Smart Cameras Using t...
“Squeezing the Last Milliwatt and Cubic Millimeter from Smart Cameras Using t...
 
"Maximize Your AI Compatibility with Flexible Pre- and Post-processing," a Pr...
"Maximize Your AI Compatibility with Flexible Pre- and Post-processing," a Pr..."Maximize Your AI Compatibility with Flexible Pre- and Post-processing," a Pr...
"Maximize Your AI Compatibility with Flexible Pre- and Post-processing," a Pr...
 
“Addressing Tomorrow’s Sensor Fusion and Processing Needs with Cadence’s Newe...
“Addressing Tomorrow’s Sensor Fusion and Processing Needs with Cadence’s Newe...“Addressing Tomorrow’s Sensor Fusion and Processing Needs with Cadence’s Newe...
“Addressing Tomorrow’s Sensor Fusion and Processing Needs with Cadence’s Newe...
 
“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...
 

Recently uploaded

Harnessing the Power of NLP and Knowledge Graphs for Opioid Research
Harnessing the Power of NLP and Knowledge Graphs for Opioid ResearchHarnessing the Power of NLP and Knowledge Graphs for Opioid Research
Harnessing the Power of NLP and Knowledge Graphs for Opioid Research
Neo4j
 
Containers & AI - Beauty and the Beast!?!
Containers & AI - Beauty and the Beast!?!Containers & AI - Beauty and the Beast!?!
Containers & AI - Beauty and the Beast!?!
Tobias Schneck
 
Session 1 - Intro to Robotic Process Automation.pdf
Session 1 - Intro to Robotic Process Automation.pdfSession 1 - Intro to Robotic Process Automation.pdf
Session 1 - Intro to Robotic Process Automation.pdf
UiPathCommunity
 
"Frontline Battles with DDoS: Best practices and Lessons Learned", Igor Ivaniuk
"Frontline Battles with DDoS: Best practices and Lessons Learned",  Igor Ivaniuk"Frontline Battles with DDoS: Best practices and Lessons Learned",  Igor Ivaniuk
"Frontline Battles with DDoS: Best practices and Lessons Learned", Igor Ivaniuk
Fwdays
 
Apps Break Data
Apps Break DataApps Break Data
Apps Break Data
Ivo Velitchkov
 
Northern Engraving | Modern Metal Trim, Nameplates and Appliance Panels
Northern Engraving | Modern Metal Trim, Nameplates and Appliance PanelsNorthern Engraving | Modern Metal Trim, Nameplates and Appliance Panels
Northern Engraving | Modern Metal Trim, Nameplates and Appliance Panels
Northern Engraving
 
Leveraging the Graph for Clinical Trials and Standards
Leveraging the Graph for Clinical Trials and StandardsLeveraging the Graph for Clinical Trials and Standards
Leveraging the Graph for Clinical Trials and Standards
Neo4j
 
GraphRAG for LifeSciences Hands-On with the Clinical Knowledge Graph
GraphRAG for LifeSciences Hands-On with the Clinical Knowledge GraphGraphRAG for LifeSciences Hands-On with the Clinical Knowledge Graph
GraphRAG for LifeSciences Hands-On with the Clinical Knowledge Graph
Neo4j
 
"NATO Hackathon Winner: AI-Powered Drug Search", Taras Kloba
"NATO Hackathon Winner: AI-Powered Drug Search",  Taras Kloba"NATO Hackathon Winner: AI-Powered Drug Search",  Taras Kloba
"NATO Hackathon Winner: AI-Powered Drug Search", Taras Kloba
Fwdays
 
"What does it really mean for your system to be available, or how to define w...
"What does it really mean for your system to be available, or how to define w..."What does it really mean for your system to be available, or how to define w...
"What does it really mean for your system to be available, or how to define w...
Fwdays
 
Y-Combinator seed pitch deck template PP
Y-Combinator seed pitch deck template PPY-Combinator seed pitch deck template PP
Y-Combinator seed pitch deck template PP
c5vrf27qcz
 
Christine's Product Research Presentation.pptx
Christine's Product Research Presentation.pptxChristine's Product Research Presentation.pptx
Christine's Product Research Presentation.pptx
christinelarrosa
 
Must Know Postgres Extension for DBA and Developer during Migration
Must Know Postgres Extension for DBA and Developer during MigrationMust Know Postgres Extension for DBA and Developer during Migration
Must Know Postgres Extension for DBA and Developer during Migration
Mydbops
 
LF Energy Webinar: Carbon Data Specifications: Mechanisms to Improve Data Acc...
LF Energy Webinar: Carbon Data Specifications: Mechanisms to Improve Data Acc...LF Energy Webinar: Carbon Data Specifications: Mechanisms to Improve Data Acc...
LF Energy Webinar: Carbon Data Specifications: Mechanisms to Improve Data Acc...
DanBrown980551
 
Poznań ACE event - 19.06.2024 Team 24 Wrapup slidedeck
Poznań ACE event - 19.06.2024 Team 24 Wrapup slidedeckPoznań ACE event - 19.06.2024 Team 24 Wrapup slidedeck
Poznań ACE event - 19.06.2024 Team 24 Wrapup slidedeck
FilipTomaszewski5
 
GNSS spoofing via SDR (Criptored Talks 2024)
GNSS spoofing via SDR (Criptored Talks 2024)GNSS spoofing via SDR (Criptored Talks 2024)
GNSS spoofing via SDR (Criptored Talks 2024)
Javier Junquera
 
GlobalLogic Java Community Webinar #18 “How to Improve Web Application Perfor...
GlobalLogic Java Community Webinar #18 “How to Improve Web Application Perfor...GlobalLogic Java Community Webinar #18 “How to Improve Web Application Perfor...
GlobalLogic Java Community Webinar #18 “How to Improve Web Application Perfor...
GlobalLogic Ukraine
 
Demystifying Knowledge Management through Storytelling
Demystifying Knowledge Management through StorytellingDemystifying Knowledge Management through Storytelling
Demystifying Knowledge Management through Storytelling
Enterprise Knowledge
 
Mutation Testing for Task-Oriented Chatbots
Mutation Testing for Task-Oriented ChatbotsMutation Testing for Task-Oriented Chatbots
Mutation Testing for Task-Oriented Chatbots
Pablo Gómez Abajo
 
From Natural Language to Structured Solr Queries using LLMs
From Natural Language to Structured Solr Queries using LLMsFrom Natural Language to Structured Solr Queries using LLMs
From Natural Language to Structured Solr Queries using LLMs
Sease
 

Recently uploaded (20)

Harnessing the Power of NLP and Knowledge Graphs for Opioid Research
Harnessing the Power of NLP and Knowledge Graphs for Opioid ResearchHarnessing the Power of NLP and Knowledge Graphs for Opioid Research
Harnessing the Power of NLP and Knowledge Graphs for Opioid Research
 
Containers & AI - Beauty and the Beast!?!
Containers & AI - Beauty and the Beast!?!Containers & AI - Beauty and the Beast!?!
Containers & AI - Beauty and the Beast!?!
 
Session 1 - Intro to Robotic Process Automation.pdf
Session 1 - Intro to Robotic Process Automation.pdfSession 1 - Intro to Robotic Process Automation.pdf
Session 1 - Intro to Robotic Process Automation.pdf
 
"Frontline Battles with DDoS: Best practices and Lessons Learned", Igor Ivaniuk
"Frontline Battles with DDoS: Best practices and Lessons Learned",  Igor Ivaniuk"Frontline Battles with DDoS: Best practices and Lessons Learned",  Igor Ivaniuk
"Frontline Battles with DDoS: Best practices and Lessons Learned", Igor Ivaniuk
 
Apps Break Data
Apps Break DataApps Break Data
Apps Break Data
 
Northern Engraving | Modern Metal Trim, Nameplates and Appliance Panels
Northern Engraving | Modern Metal Trim, Nameplates and Appliance PanelsNorthern Engraving | Modern Metal Trim, Nameplates and Appliance Panels
Northern Engraving | Modern Metal Trim, Nameplates and Appliance Panels
 
Leveraging the Graph for Clinical Trials and Standards
Leveraging the Graph for Clinical Trials and StandardsLeveraging the Graph for Clinical Trials and Standards
Leveraging the Graph for Clinical Trials and Standards
 
GraphRAG for LifeSciences Hands-On with the Clinical Knowledge Graph
GraphRAG for LifeSciences Hands-On with the Clinical Knowledge GraphGraphRAG for LifeSciences Hands-On with the Clinical Knowledge Graph
GraphRAG for LifeSciences Hands-On with the Clinical Knowledge Graph
 
"NATO Hackathon Winner: AI-Powered Drug Search", Taras Kloba
"NATO Hackathon Winner: AI-Powered Drug Search",  Taras Kloba"NATO Hackathon Winner: AI-Powered Drug Search",  Taras Kloba
"NATO Hackathon Winner: AI-Powered Drug Search", Taras Kloba
 
"What does it really mean for your system to be available, or how to define w...
"What does it really mean for your system to be available, or how to define w..."What does it really mean for your system to be available, or how to define w...
"What does it really mean for your system to be available, or how to define w...
 
Y-Combinator seed pitch deck template PP
Y-Combinator seed pitch deck template PPY-Combinator seed pitch deck template PP
Y-Combinator seed pitch deck template PP
 
Christine's Product Research Presentation.pptx
Christine's Product Research Presentation.pptxChristine's Product Research Presentation.pptx
Christine's Product Research Presentation.pptx
 
Must Know Postgres Extension for DBA and Developer during Migration
Must Know Postgres Extension for DBA and Developer during MigrationMust Know Postgres Extension for DBA and Developer during Migration
Must Know Postgres Extension for DBA and Developer during Migration
 
LF Energy Webinar: Carbon Data Specifications: Mechanisms to Improve Data Acc...
LF Energy Webinar: Carbon Data Specifications: Mechanisms to Improve Data Acc...LF Energy Webinar: Carbon Data Specifications: Mechanisms to Improve Data Acc...
LF Energy Webinar: Carbon Data Specifications: Mechanisms to Improve Data Acc...
 
Poznań ACE event - 19.06.2024 Team 24 Wrapup slidedeck
Poznań ACE event - 19.06.2024 Team 24 Wrapup slidedeckPoznań ACE event - 19.06.2024 Team 24 Wrapup slidedeck
Poznań ACE event - 19.06.2024 Team 24 Wrapup slidedeck
 
GNSS spoofing via SDR (Criptored Talks 2024)
GNSS spoofing via SDR (Criptored Talks 2024)GNSS spoofing via SDR (Criptored Talks 2024)
GNSS spoofing via SDR (Criptored Talks 2024)
 
GlobalLogic Java Community Webinar #18 “How to Improve Web Application Perfor...
GlobalLogic Java Community Webinar #18 “How to Improve Web Application Perfor...GlobalLogic Java Community Webinar #18 “How to Improve Web Application Perfor...
GlobalLogic Java Community Webinar #18 “How to Improve Web Application Perfor...
 
Demystifying Knowledge Management through Storytelling
Demystifying Knowledge Management through StorytellingDemystifying Knowledge Management through Storytelling
Demystifying Knowledge Management through Storytelling
 
Mutation Testing for Task-Oriented Chatbots
Mutation Testing for Task-Oriented ChatbotsMutation Testing for Task-Oriented Chatbots
Mutation Testing for Task-Oriented Chatbots
 
From Natural Language to Structured Solr Queries using LLMs
From Natural Language to Structured Solr Queries using LLMsFrom Natural Language to Structured Solr Queries using LLMs
From Natural Language to Structured Solr Queries using LLMs
 

"Methods for Creating Efficient Convolutional Neural Networks," a Presentation from Xnor.ai

  • 1. © 2019 xnor.ai Methods for Creating Efficient Convolutional Neural Networks Mohammad Rastegari xnor.ai May 2019
  • 2. © 2019 xnor.ai Approaches to Efficient CNN • Model design optimization • Lower Precision (Quantization) • Binary • Sparse Models • Lookup based CNN • Compact Network Design • Elastic model, Hierarchical convolution, Dimension-wise convolution • Training Optimization • Label Refinery
  • 3. © 2019 xnor.ai … … Convolutional Neural Networks
  • 4. © 2019 xnor.ai • 1B – 20B FLOPs Number of Operations : • 0.25 – 3 fps Inference time on CPU : GPU ! * + − ×
  • 5. © 2019 xnor.ai Approaches to Efficient CNN • Model design optimization • Lower Precision (Quantization) • Binary • Sparse Models • Lookup based CNN • Compact Network Design • Elastic model, Hierarchical convolution, Dimension-wise convolution • Training Optimization • Label Refinery
  • 6. © 2019 xnor.ai Lower Precision 32-bit 1-bit Reducing Precision • Saving Memory • Saving Computation {-1,+1} {0,1} MUL XNOR ADD, SUB Bit-Count (popcount) 8-bit
  • 7. © 2019 xnor.ai Why Binary? Binary Instructions • AND, OR, XOR, XNOR, PoPCount (Bit-Count) Low Power Device Easy to Implement in hardware
  • 8. © 2019 xnor.ai + − × 1x 1x Operations Memory Computation + − ~32x ~2x XNOR Bit- count ~32x ~58x Binary Weight Networks XNOR-Networks Theoretical Improvement * * * *
  • 9. © 2019 xnor.ai filter k 2 R , K = A ⇤k, where 8ij ki j = w ⇥h . K or all sub-tensors in the input I . K i j corresponds to β for e location ij (across width and height). This procedure is ure2. Once weobtained thescaling factor ↵ for theweight n I (denoted by K ), we can approximate the convolution filter W mainly using binary operations: W ⇡ (sign(I ) ~ sign(W )) K ↵ (11) s the procedure explained in section 3.2 for approximating a convo- ns. onal vector where all of its enteries are 1. 1T can be factored and theoptimal solutions can beachieved from equation 2 as sign(Y ) = sign(X T ) sign(W ) = H ⇤T B ⇤ (9) ependent, knowing that Y i = X i W i then, = E [|X i |] E [|W i |] therefore, P |X i ||W i | n ⇡ ✓ 1 n kX k` 1 ◆✓ 1 n kW k`1 ◆ = β⇤ ↵⇤ (10) we convolve A with a 2D filter k 2 R , K = A ⇤k, where 8ij ki j = w ⇥h contains scaling factors β for all sub-tensors in the input I . K i j corresponds to β a sub-tensor centered at the location ij (across width and height). This procedur shown in thethird row of figure2. Once weobtained thescaling factor ↵ for thewe and β for all sub-tensors in I (denoted by K ), we can approximate the convolu between input I and weight filter W mainly using binary operations: I ⇤W ⇡ (sign(I ) ~ sign(W )) K ↵ lution using binary operations. where 1 is an n-dimensional vector where all of its enteries are 1. 1T ca out from theoptimization and theoptimal solutions can beachieved from follow C⇤ = sign(Y ) = sign(X T ) sign(W ) = H ⇤T B ⇤ Since |X i |, |W i | are independent, knowing that Y i = X i W i then, E [|Y i |] = E [|X i ||W i |] = E [|X i |] E [|W i |] therefore, γ⇤ = P |Y i | n = P |X i ||W i | n ⇡ ✓ 1 n kX k` 1 ◆✓ 1 n kW k`1 ◆ = β c⇥w⇥h WB WB WB = sign(W) *
  • 10. © 2019 xnor.ai Quantization Error WB = sign(W) _ 0.75 WB
  • 11. © 2019 xnor.ai Optimal Scaling Factor WB
  • 12. © 2019 xnor.ai Binary Input and Binary Weight (XNOR-Net) a sub-tensor centered at the location ij (across width and height). This proced shown in thethird row of figure2. Once weobtained thescaling factor ↵ for thew and β for all sub-tensors in I (denoted by K ), we can approximate the convo between input I and weight filter W mainly using binary operations: I ⇤W ⇡ (sign(I ) ~ sign(W )) K ↵ i j location ij (across width and height). This procedure is re2. Once weobtained thescaling factor ↵ for theweight I (denoted by K ), we can approximate the convolution lter W mainly using binary operations: ⇡ (sign(I ) ~ sign(W )) K ↵ (11) nal vector where all of its enteries are 1. 1T can be factored and theoptimal solutions can beachieved from equation 2 as gn(Y ) = sign(X T ) sign(W ) = H ⇤T B ⇤ (9) endent, knowing that Y i = X i W i then, E [|X i |] E [|W i |] therefore, P |X i ||W i | n ⇡ ✓ 1 n kX k` 1 ◆✓ 1 n kW k`1 ◆ = β⇤ ↵⇤ (10) volving weight filter W 2 Rc⇥w⇥h (wherewi n w, hi n WBXB
  • 13. © 2019 xnor.ai How to train a CNN with binary filters?
  • 14. © 2019 xnor.ai Training Binary Weight Networks Naive Solution: 1. Train a network with real value parameters 2. Binarize the weight filters
  • 15. © 2019 xnor.ai 0 10 20 30 40 50 60 70 80 Top-1 (%) ILSVRC2012 Full Precision
  • 16. © 2019 xnor.ai . . . . . .W . . . . . .WB Binarization
  • 17. © 2019 xnor.ai . . . . . . Person Dog . . . . . .W Binarization
  • 18. © 2019 xnor.ai Training XnorNet W = W - ηGw . . . . . . . . . . . . . . . . . . Gw W Train for binary weights: [XNOR-Networks, Rastegari et al, ECCV2016]
  • 19. © 2019 xnor.ai 0 10 20 30 40 50 60 70 80 Top-1 (%) ILSVRC2012 0.2 [XNOR-Networks, Rastegari et al, ECCV2016]
  • 20. © 2019 xnor.ai Top-1 (%) ILSVRC2012 0.2 0 10 20 30 40 50 60 70 80
  • 21. © 2019 xnor.ai Approaches to Efficient CNN • Model design optimization • Lower Precision (Quantization) • Binary (XNOR-Net) • Sparse Models • Lookup based CNN • Compact Network Design • Elastic model, Hierarchical convolution, Dimension-wise convolution • Training Optimization • Label Refinery
  • 22. © 2019 xnor.ai … … Lookup Based CNN
  • 26. © 2019 xnor.ai How to train the discrete indexing?!!!! *
  • 29. © 2019 xnor.ai 0 10 20 30 40 Speed-up AccuracyRate Image Classification Few-shot Training Few Iteration Training OnDeviceTraining
  • 30. © 2019 xnor.ai Approaches to Efficient CNN • Model design optimization • Lower Precision (Quantization) • Binary (XNOR-Net) • Sparse Models • Lookup based CNN • Compact Network Design • Elastic model, Hierarchical convolution, Dimension-wise convolution • Training Optimization • Label Refinery
  • 31. © 2019 xnor.ai Elastic: Instance Specific Efficiency
  • 32. © 2019 xnor.ai Challenging vs. Simple Images
  • 33. © 2019 xnor.ai Challenging vs. Simple Images
  • 34. © 2019 xnor.ai Challenging vs. Simple Images
  • 35. © 2019 xnor.ai Challenging vs. Simple Images
  • 36. © 2019 xnor.ai Challenging vs. Simple Images
  • 37. © 2019 xnor.ai ng Elas- evaluate eshow ermore, ransfer classifi- Xt [34], be aug- rch Im- but no tandard Figure4: Imagenet Accuracy vs. FLOPSand ParametersThis figure shows our Elastic model can achieve a lower error without any extra (or with lower) computational cost. parameters. Table2comparesthetop-1 andtop-5 error rates of all of the base models with the Elastic augmentation (in-
  • 38. © 2019 xnor.ai Approaches to Efficient CNN • Model design optimization • Lower Precision (Quantization) • Binary (XNOR-Net) • Sparse Models • Lookup based CNN • Compact Network Design • Elastic model, Hierarchical convolution, Dimension-wise convolution • Training Optimization • Label Refinery
  • 39. © 2019 xnor.ai Standard Convolution Group Convolution Depth-wise Convolution Efficient Convolution * * * *
  • 40. © 2019 xnor.ai Standard Convolution Dilated Convolution
  • 41. © 2019 xnor.ai Depth-wise Dilated Convolution (DDConv)
  • 43. © 2019 xnor.ai Mehtaet al. (a) RGB without HFF with HFF (b) Object Boundary Detection Gridding Effect Standard Block structure
  • 45. © 2019 xnor.ai Mehtaet al. (a) RGB without HFF with HFF (b)Gridding Effect Standard Block structure No Gridding Effect Hierarchical structure Object Boundary Detection
  • 46. © 2019 xnor.ai Model FLOPs mIOU HDDConv 1.4 B 69.1 DeepLabV3 2.84 B 71.8 Semantic Object Segmentation
  • 47. © 2019 xnor.ai Approaches to Efficient CNN • Model design optimization • Lower Precision (Quantization) • Binary (XNOR-Net) • Sparse Models • Lookup based CNN • Compact Network Design • Elastic model, Hierarchical convolution, Dimension-wise convolution • Training Optimization • Label Refinery
  • 48. © 2019 xnor.ai Tensor Encoding Capacity *
  • 49. © 2019 xnor.ai Tensor Encoding Capacity *
  • 50. © 2019 xnor.ai Tensor Encoding Capacity Ordering in channels has not being used for encoding *
  • 51. © 2019 xnor.ai Tensor Encoding Capacity Expensive Channel Fusion. O(c^2.h.w)Depth-Conv. O(c.h.w.k^2) * *
  • 52. © 2019 xnor.ai Dimension-wise Convolution Dim-Conv. O(c.h.w.k^2) c h w * * *
  • 53. © 2019 xnor.ai Dimension-wise Convolution Dim-Conv. O(c.h.w.k^2) contains scaling factors β for all sub-tens a sub-tensor centered at the location ij ( shown in thethird row of figure2. Oncew and β for all sub-tensors in I (denoted b between input I and weight filter W main I ⇤W ⇡ (sign(I ) Efficient Channel Fusion. O(c^2+h.w) contains scaling factors β for all sub-tensors in the inp a sub-tensor centered at the location ij (across width shown in thethird row of figure2. Onceweobtained the and β for all sub-tensors in I (denoted by K ), we can between input I and weight filter W mainly using binar I ⇤W ⇡ (sign(I ) ~ sign(W )) * * *
  • 55. © 2019 xnor.ai FLOPs vs. Accuracy on Image Classification ResNet-50 XNOR-res50 HDDconv HDDConv HDDConv LCNN LCNN FLOPS(Log Millions) ELASTIC HDDConvDimConv HDDConv DimConv HDDConv DimConv Accuracy
  • 56. © 2019 xnor.ai Approaches to Efficient CNN • Model design optimization • Lower Precision (Quantization) • Binary (XNOR-Net) • Sparse Models • Lookup based CNN • Compact Network Design • Elastic model, Hierarchical convolution, Dimension-wise convolution • Training Optimization • Label Refinery
  • 57. © 2019 xnor.ai Components in a Supervised Learning System Data • ImageNet, MSCOCO, SUN, … • Data Augmentations Model • SVM, CNN • Optimization Techniques (SGD,ADAM, RMSProp,…) Label • ?!!
  • 58. © 2019 xnor.ai Challenges with current labeling paradigm Persian Cat ball • Incomplete
  • 59. © 2019 xnor.ai Challenges with current labeling paradigm Random cropping
  • 60. © 2019 xnor.ai Challenges with current labeling paradigm • Inconsistent Dough Butternut Squash
  • 61. © 2019 xnor.ai Challenges with current labeling paradigm • Inconsistent Dough Butternut Squash
  • 62. © 2019 xnor.ai Challenges with current labeling paradigm chrysanthemum dog silky terrier Car mirror Sameamountofpenalization • Taxonomy dependency
  • 63. © 2019 xnor.ai Labels should be: Soft Informative Dynamic Cat → 80% Ball → 20% Dog --> 60% Cat --> 30% Bear --> 10% Dog --> 60% Cat --> 10% Bear --> 30% Cat → 1 % Ball → 99%
  • 64. © 2019 xnor.ai Label Refinery Ground- truth Label Data burrito burrito plate eggnog burrito plate restaurant Refinery Top-1: 57.93 Top-1: 59.97 Top-1: 60.87 Top-1: 61.22 burrito plate restaurant Refined Label Data Refinery Refined Label Data Refinery Refined Label Data Model
  • 65. © 2019 xnor.ai Label Refinery Ground- truth Label Data burrito burrito plate eggnog burrito plate restaurant Refinery Top-1: 57.93 Top-1: 59.97 Top-1: 60.87 Top-1: 61.22 burrito plate restaurant Refined Label Data Refinery Refined Label Data Refinery Refined Label Data Model
  • 66. © 2019 xnor.ai Label Refinery Ground- truth Label Data burrito burrito plate eggnog burrito plate restaurant Refinery Top-1: 57.93 Top-1: 59.97 Top-1: 60.87 Top-1: 61.22 burrito plate restaurant Refined Label Data Refinery Refined Label Data Refinery Refined Label Data Model
  • 67. © 2019 xnor.ai Model Top-1 Top-5 Top-1 Top-5 AlexNet 57.93 79.41 66.28 86.13 MobileNet-1 68.53 88.14 73.39 91.07 MobileNet-0.75 65.93 86.28 70.92 89.68 MobileNet-0.5 63.03 84.55 66.66 87.07 MobileNet-0.25 50.65 74.42 54.62 77.92 ResNet-50 75.7 92.81 76.5 93.12 ResNet-34 73.39 91.32 75.06 92.35 ResNet-18 69.7 89.26 72.52 90.73 ResNetXnor-50 63.1 83.61 73.31 89.18 VGG16 70.1 88.54 75 92.22 VGG19 71.39 89.44 75.46 92.52 DarkNet19 70.6 89.13 74.47 91.94 Label RefineryStandard Training © 2019 xnor.ai Model Top-1 Top-5 Top-1 Top-5 AlexNet 57.93 79.41 66.28 86.13 MobileNet-1 68.53 88.14 73.39 91.07 MobileNet-0.75 65.93 86.28 70.92 89.68 MobileNet-0.5 63.03 84.55 66.66 87.07 MobileNet-0.25 50.65 74.42 54.62 77.92 ResNet-50 75.7 92.81 76.5 93.12 ResNet-34 73.39 91.32 75.06 92.35 ResNet-18 69.7 89.26 72.52 90.73 ResNetXnor-50 63.1 83.61 73.31 89.18 VGG16 70.1 88.54 75 92.22 VGG19 71.39 89.44 75.46 92.52 DarkNet19 70.6 89.13 74.47 91.94 Label RefineryStandard Training
  • 68. © 2019 xnor.ai 74 How far we can get with this efficiency? Server & CloudEdge & Embedded Devices Mobile Devices Compute Capability & Price HighLow Traditional Home of AI FPGA AI/Neural Accelerator GPU CPUs
  • 69. © 2019 xnor.ai Thank You !!! 1. ELASTIC: Improving CNNs with Instance Specific Scaling PoliciesH Wang, A Kembhavi, A Farhadi, A Yuille, M Rastegari arXiv preprint arXiv:1812.05262 2. ESPNetv2: A Light-weight, Power Efficient, and General Purpose Convolutional Neural NetworkS Mehta, M Rastegari, L Shapiro, H Hajishirzi arXiv preprint arXiv:1811.11431 3. Label refinery: Improving imagenet classification through label progressionH Bagherinezhad, M Horton, M Rastegari, A Farhadi arXiv preprint arXiv:1805.02641 4. Espnet: Efficient spatial pyramid of dilated convolutions for semantic segmentationS Mehta, M Rastegari, A Caspi, L Shapiro, H Hajishirzi . Proceedings of the European Conference on Computer Vision (ECCV), 552-568 5. Xnor-net: Imagenet classification using binary convolutional neural networksM Rastegari, V Ordonez, J Redmon, A Farhadi European Conference on Computer Vision, 525-542 6. Lcnn: Lookup-based convolutional neural networkH Bagherinezhad, M Rastegari, A Farhadi. Proceedings of the IEEE Conference on Computer Vision and Pattern …