Layer Number
Layer Size
Volumetric layer and filter
D L S T M A U T O E N C O D E R
Layer dimensions
To be first presented at HPE Discover 2018 in Las Vegas and then presented to audiences globally, the content developed needed to live in both digital and physical format seamlessly. The scope of development considered all steps of the final deployment process, structuring a visual logic and the presentation cadence of the design artifacts to allow for both a guided tour in a touch-screen enabled presentation booth,
and to live as a stand-alone unit used to educate the industry remotely.
3D Layer Volume
Block Layer Numbers
After many rounds of review and refinement with machine learning and neural network experts, the results of the visualization were complete and the industry flavoring accurate to the topics intended to be covered during partner conversations. The deployment of the experience was piloted by the deployment of the experience at the Hewlett Packard Labs team booth at HPE Discover 2018 in Las Vegas, before its deployment to online channels.
Input: 50 data points (neurons)
5 hidden recurrent layers, 128 neurons each
Output: 50 neurons
Non-volumetric layer and filter
S T E P 4
V I S U A L I Z E D N E T W O R K S
The A.R. city serves as a landing point for the user to begin the interaction with the virtual world, displaying a city that contains the 5 industry cases covered by the content, embodied in meaningful building shapes within an augmented reality world.
The orientation of each building is strategic to provoke the best interface engagement, with each of the five buildings within the A.R. model facing different directions of the physical table where they first appear.
The user is first instructed by the printed marker on the tabletop to scan it
using the iPad, unleashing the A.R. experience.
S T E P 5
I N D U S T R Y V E R T I C A L F L A V O R I N G
Filter Dimensions
After completing the visualization and logical proofing of the visual layout using the knowledge of industry experts, the designs were to serve as conversation topics embedded in relevant and relatable industry vertical stories. The five elected verticals were the potential most auspicious users of the neural network computing infrastructure currently being developed and supported by Hewlett Packard.
S C R E E n 3
O P T I M A L H A R D W A R E
v o l u m e t r i c l a y e r s & f i l t e r s
A U T O M A T E D M E E T I N G N O T E S
Filter dimensions
S T E P 2
D E V E L O P A D E S I G N K E Y
Layer 3: 32x32x8 (64 3x3x3 filters)
S T E P 1
I N T E R P R E T F O R M U L A L O G I C
Layer 33 : 1x6, fully connected, 6 neurons
N O N - V O L U M E T R I C L A Y E R S & F I L T E R S
A U T O N O M O U S V E H I C L E
D E E P L E A R N I N G C O O K B O O K
H P E D I S C O V E R 2 0 1 8
Layered Repeats
ResNetB3D:
Layer 0 (input data): volume 64x64x16
Layer 1: 64x64x16 x 48 (after processing with 48 filters 3x3x3)
Layer 2: 32x32x8 x 64 (64 3x3x3 filters, stride 2)
Layer 3: 32x32x8 (64 3x3x3 filters)
Layers 4-5: repeat 2 more times layer 3
Layer 6: 32x32x8 x 64 (64 1x1x1 filters)
Layer 7: 16x16x4 x 64 (64 3x3x3 filters, stride 2)
Layer 8: 16x16x4 x 128 (128 1x1x1 filters)
Layer 9: 16x16x4 x 64 (64 1x1x1 filters)
Layer 10: 16x16x4 x 64 (64 3x3x3 filters)
Layer 11: 16x16x4 x 128 (128 1x1x1 filters)
Layers 12-17: repeat 2 more times layers 9,10,11
Layer 18: 16x16x4 x 128 (128 1x1x1 filters)
Layer 19: 8x8x2 x 128 (128 3x3x3 filters, stride 2)
Layer 20: 8x8x2 x 256 (256 1x1x1 filters)
Layer 21: 8x8x2 x 128 (128 1x1x1 filters)
Layer 22: 8x8x2 x 128 (128 3x3x3 filters)
Layer 23: 8x8x2 x 256 (256 1x1x1 filters)
Layers 24-32: repeat 3 more times layers 21, 22, 23
Layer 33: 1x6, fully connected, 6 neurons
Layer Type
WEBSITE AND CONTENT DESIGNED BY ©EDUARDO M NAUIACK 2020 ALL RIGHTS RESERVED
Layer 0 (input data): image 224x224x3 (width x height x color channel)
Layer 1: 112x112x64 (after processing with 64 filters 7x7 each, stride 2)
Layer 2: 56x56x64 (64 1x1 filters)
Layer 3: 56x56x64 (64 3x3 filters)
Layer 4: 56x56x256 (256 1x1 filters)
Layers 5-10: Repeat 2 more times layers 2, 3, 4:
Layer 5: 56x56x64 (64 1x1 filters)
Layer 6: 56x56x64 (64 3x3 filters)
Layer 7: 56x56x256 (256 1x1 filters)
Layer 8: 56x56x64 (64 1x1 filters)
Layer 9: 56x56x64 (64 3x3 filters)
Layer 10: 56x56x256 (256 1x1 filters)
Layer 11: 28x28x128 (128 1x1 filters)
Layer 12: 28x28x128 (128 3x3 filters)
Layer 13: 28x28x512 (512 1x1 filters)
Layers 14-34: repeat 7 more times layers 11, 12, 13
Layer 35: 14x14x256 (256 1x1 filters)
Layer 36: 14x14x256 (256 3x3 filters)
Layer 37: 14x14x1024 (1024 1x1 filters)
Layers 38-142: repeat 35 more times layers 35, 36, 37
Layer 143: 7x7x512 (512 1x1 filters)
Layer 144: 7x7x512 (512 3x3 filters)
Layer 145: 7x7x2048 (2048 1x1 filters)
Layers 146-151: repeat 2 more times layers 143, 144, 145Layer 152: 1x1000 - a fully connected with 1000 neurons
2 . 5 D A . R G R A P H I C S T R E A T M E N T
S T E P 3
V I S U A L I Z E l a y e r s & F I L T E R S
Number of filters
D E E P M N I S T
O I L A N D G A S
I M A G E G U I D E D S U R G E R Y
Layer 0 (input data): volume 64x64x16
Layer 1: 64x64x16 x 48 (after processing with 48 filters 3x3x3)
Layer 2: 32x32x8 x 64 (64 3x3x3 filters, stride 2)
Layer 3: 32x32x8 (64 3x3x3 filters)
Layers 4-5: repeat 2 more times layer 3
Layer 6: 32x32x8 x 64 (64 1x1x1 filters)
Layer 7: 16x16x4 x 64 (64 3x3x3 filters, stride 2)
Layer 8: 16x16x4 x 128 (128 1x1x1 filters)
Layer 9: 16x16x4 x 64 (64 1x1x1 filters)
Layer 10: 16x16x4 x 64 (64 3x3x3 filters)
Layer 11: 16x16x4 x 128 (128 1x1x1 filters)
Layers 12-17: repeat 2 more times layers 9,10,11
Layer 18: 16x16x4 x 128 (128 1x1x1 filters)
Layer 19: 8x8x2 x 128 (128 3x3x3 filters, stride 2)
Layer 20: 8x8x2 x 256 (256 1x1x1 filters)
Layer 21: 8x8x2 x 128 (128 1x1x1 filters)
Layer 22: 8x8x2 x 128 (128 3x3x3 filters)
Layer 23: 8x8x2 x 256 (256 1x1x1 filters)
Layers 24-32: repeat 3 more times layers 21, 22, 23
Layer 33: 1x6, fully connected, 6 neurons
S C R E E n 2
U S E C A S E M O D E L
D I S P L A Y T O U C H S C R E E N S
The challenge consisted of interpreting the mathematical formula of diverse neural network designs used in industry relevant verticals, developing a design logic to visualize them, and make the content graspable and memorable for global customers and partners.
3D Filters
Block Number
Hewlett Packard Labs is the division of HPE responsible for the most advanced, cutting edge and relevant advancements supporting the future of computing infrastructure. Within this team of scientists and engineering gurus, machine learning and neural networks play a large role in the design of computational frameworks and the hardware to run the world of tomorrow smoothly. Hewlett Packard Labs prepared for the necessary conversations with industry leaders and global partners present in global events by partnering with Aniden Interactive in designing a more understandable diagram of neural networks design.
H E W L E T T P A C K A R D L A B S
N E U R A L N E T V I S U A L I Z A T I O N
CREATIVE dir. + DESIGN + DEPLOYMENT
Layers 12-17: repeat 2 more times layers 9,10,11
Layer Type
Filter
Layer blocks, indicate layer repeats and stride partioning
HPE DISCOVER 2018 – LAS VEGAS, NV / GLOBAL PARTNERS
R E S N E T 1 5 2
Layer Size
Deployment sample
5760px by 1080px
P R E S C R I P T I V E M A I N T E N A N C E
S T E P 5
D E P L O Y M E N T
The experience lived in a physical wall developed to ground the narrative onto future cities using the HPE style iconography, the landscape was designed in Illustrator and printed in high resolution along with 3 x 50" touch-enabled screens where the content was displayed and navigated by the booth hosts. The touch navigation experience
was deployed in Powerpoint format utilizing invisible block click-event rigging.
Input size (input layer): 10000
6 hidden fully-connected layers, 512 neurons each
Output layer: 20x20x20 grid (8000 neurons)
The experience was immensely successful at attracting guests and starting relevant conversations between industry experts, pushed to online mediums after the successful physical deployment pilot at HPE Discover.
Special thanks to Jessica Visser, Natalia Vassilieva and Sergey Serebryakov.
R E S N E T B 3 D
S C R E E n 1
U S E C A S E D E E P D I V E
Number of Layered Repeats
D E E P S P E E C H
Main Menu
An Evolving World Journey deployed using an iPad, the Artificial Reality city world leads the user interaction towards the touch-point of each industry vertical being supported by HPE A.I. technologies.
The environment of each selected (five) industry segment is custom to the line of storytelling leading to the HPE Deep Learning Cookbook content. The visual style is heavily guided by HPE storytelling iconography brought into 3D A.R. space.
D E E P L E A R N I N G C O O K B O O K
A U G M E N T E D R E A L I T Y E X P E R I E N C E