Speaker: Barnam Bora, Head of AI/ML, APAC, AWS
Customer Speaker: Guangda Li, Co-founder & CTO, ViSenze
Note: This is part 2 of the deck.
WS offers different paths for building and deploying scalable ML solutions. This session provides an insight to how AWS customers are building intelligent systems powered by AI and ML. Learn how these services, in conjunction with the large number of complementary AWS technologies, provide a great platform for our customers to build their own AI and ML powered solutions and drive business value. Towards the latter part of this session, hear how customers are deploying their ML on AWS and can now leverage Marketplace to monetise their models.
15. Visual Search and Recognition at ViSenze
Guangda LI
CTO and Co-founder
www.visenze.com
16. • Started as Spin off from NExT(NUS-Tsinghua Research centre); Raised 34M USD funding from VCs
• Based on computer vision and deep learning research, ViSenze provides visual search and image
recognition solution for some of the largest companies in the world
• Customers include the top fast fashion retailer, sports brand and mobile phone manufacturer
ViSenze has been recognized amongst:
• Top 5 deep learning companies - VentureBeat 2017
• Top AI Product AI (Retail) – CogX London 2017
• Top 40 global Breakthrough Brands - InterBrand 2017
About ViSenze
Top 25 AI companies
in 2018
17. Under the Hood
• Built by Computer Vision scientists and software development
experts (40+ R&D in CV, ML, Infra)
• 4 patent applications (granted and pending) in various stages
• Global scale distributed architecture supporting over 1 Billion
• Low latency and high-throughput architect design
• In-house distributed GPU training platform and tool development
• Independent validations: ImageNet, client evaluations
Deep Learning & Computer Vision AI DevOps Platform
Continuous Training Domain Models
ViSenze singled out in keynote speech by
Prof Li Fei Fei of Stanford at CVPR 2017
ViSenze partners Nvidia using latest GPUs for
image processing and CNN model training
18. The Rise of Visual Content on Internet
Gen Z prefers to communicate with Images Image Growth - More than 3B photos per day
19. Major players are already driving this shift of visual search adoption
250M visual searches/month - Pinterest, May 2017
360M visual searches/month - Taobao, July 2017
https://www.forbes.com/sites/kathleenchaykowski/2017/02/08/pinterest-debuts-new-camera-lens-search-tools-to-find-real-world-objects-online/#3223bbc060e1
Amazon Visual Search Pinterest Lens Samsung Bixby Vision Google Lens
20. H&M, ASOS and UNIQLO Visual Search powered by ViSenze
Image Search is now a common feature on leading retailer apps
22. Improving User Engagement Through Artificial Intelligence
A.I. powered visual search and recognition
solutions improve engagement and
conversions
30% 50% 5x 160%
higher conversions
on image search
over text based
search
higher CTR of
shoppers who click
on visually similar
products
higher conversion
rates for shoppers
clicking on visually
similar products
increase in
engagement for
shoppers who used
find similar
23. Platform Solution: ViSenze Shopping Lens on Smartphones
200M+ Products from over 800+ major global retailer on ViSenze global affiliate network
26. Tagging Solutions and Use Cases
Automated Tagging Catalogue Management
Tag entire image libraries based on visual attributes for better
search results
Image Filtering and Moderation
Tag entire image libraries based on visual attributes
Improve discoverability and conversions
Fashion Attributes
Style Attributes
Image Quality
27. Tagging Solutions and Use Cases - Example
Fashion Attribute
Fashion style
& Occasion
28. Fine-grained Fashion Attribute Recognition
25 types of neckline:
zip_neck
henley
off_shoulder
scoop
crew_neck
round_neck
v_neck
cowl
turtleneck
...
peter_pan
keyhole
one_shoulder
tie_up_neck
other
60 types of product categories:
sweater
sweatshirt_hoodies
blouse
shirt
t_shirt
polo_shirt
top
camisole
tank_top
…
clutch_purse
card_holder
pouch
wallet
others
14 types of product pattern:
stripes
text
animal_print
big_graphic
...
plaid
solid
other
29. Life of A ML pipeline: Continuous Data Cleaning
Data Management Challenges in Production Machine Learning, Alkis Polyzotis etc 2017, SIGMOD
30. The Gap between DL Software and Production DL System
TFX: A TensorFlow-Based Production-Scale Machine Learning Platform D. Baylor etc KDD, 2017.
Deep learning framework is a
small component for production
level deep learning development
32. Hidden Facts for Production Deep learning Development
Key winning formula for DL-based applications
• Transfer domain Expertise into clear requirement
• Large scale high quality training data and data management
• Intensive and faster feedback
Continuous performance improvement
• Hard to result from methodology improvement
• Usually results from data driven improvement
33. DNN Inference Engine on AWS
● Caffe/MXNet/Caffe2 model support
● Batch inference
● TensorRT for Nvidia GPU/OpenVINO for Intel CPU
● Deployment on AWS P2/C series
● scale to 4000+ CPUs, with peak throughput 5M
images/hour.Based on the price, choose the most
cost-efficient instance type.