​

Як Deep Learning �змінить правила гри

в бізнесі.

Borys Pratsiuk, Ph.D, Head of R&D.

Who am I?

First project,�C, embedded

2004

Engineer, R&D Lab, Tescom, South Korea

2006

2013

Assistant professor, Kiev Polytechnic Institute

2012

Ph.D Solidstate Electronic

2015 - ...

Senior Android

Team Lead

Android Architect

​

2009

Android Developer

Head of R&D Engineering

​

2007

b_pratsiuk

bopr@ciklum.com

Borys Pratsiuk, �Ph.D.

Inspire brilliant minds to innovate and create.

2

Agenda

Що нового в Deep Learning а останній рік

Як великі компанії освоюють AI для автоматизації свої процесів

Питання

Nvidia Deep Learning partner

Inspire brilliant minds to innovate and create.

3

ARTIFICIAL

INTELLIGENCE

4

What is scientific research?

Scientific research is a fundamental background to test �any revolutionary business ideas

Zhang, W., Yu, Q., Siddiquie, B., Divakaran, A., & Sawhney, H. (2015). "Snap-n-Eat": food recognition and nutrition estimation on a smartphone. Journal of Diabetes Science and Technology, 9(3), 525-533. doi: 10.1177/1932296815582222

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3224860/

​

Inspire brilliant minds to innovate and create.

5

https://beamandrew.github.io/deeplearning/2017/02/23/deep_learning_101_part1.html

Inspire brilliant minds to innovate and create.

6

Inspire brilliant minds to innovate and create.

7

Deep Learning today

Healthcare

Finance

Robotics

​

Telecom

Travel

Automotive

Inspire brilliant minds to innovate and create.

8

CVPR 2018 :Trends, hypes and other cool stuff

9

140+ exhibitors and sponsors to participate.

9

10

11

Detection Beyond Bounding Boxes

12

COCO + Mapillary

13

OLD OBJECT DETECTION IS DEAD

13

Attention in Neural Networks.

https://youtu.be/Q57rzaHHO0k

Transfer Learning.

https://youtu.be/yofjFQddwHE

CVPR 2018 - http://taskonomy.stanford.edu/

Real time video tracking MDNet.

https://youtu.be/zYM7G5qd090

Як це роблять

Великі компанії?

Investment portfolio monitoring

To develop a data collection and analysis platform and automate the sourcing process for B2B leads (discovery, research, and follow-on analysis)

​

Data platform, “Ranking” algorithm, Data Access Application

​

​

• The platform enables VP to have an instant access to the consistent and validated data within days and in some cases within hours which enables them quickly to spot opportunities, quantify potential and follow changes;
• The ML algorithm automates the research process and enables analysts to look only at high potential opportunities;
• The Chrome extension helps analysts stay concentrated, increases efficiency and saves time up to 1h per day.
• Analytical platform identifies significant changes of every company and bring the attention of the analyst to these changes.
• Monitor 6 million companies simultaneously

​

​

18

Skills | Knowledge | Collaboration

Data preparation for Scientists

At the time of the project, Axonix had more than 2.5 Pb of raw real-time bidding (RTB) logs that contained anonymised information about user behaviour. To minimize processing costs, the proposed data modelling pipeline required an incremental approach to update models every week without rerunning the pipeline with historical data.

​

Data processing pipeline efficiently aggregated RTB records, anomaly detection algorithm, developed modelling pipeline has model tuning, feature selection, model evaluation and reporting capabilities.

​

​

Axonix achieved very significant savings by replacing the previous solution with this in house system designed by Ciklum. The payback time for the project investment was measured in months rather than years.

​

​

19

Skills | Knowledge | Collaboration

Demographics Prediction

Company that operates within video related analytics sector. There was high interest from video creators in the user distribution that watch their content. Company also was interested in using this data for more concise and intelligent advertising targeting purposes.

​

Two methods of solution were proposed each with its strengths and disadvantages but similar accuracy. 

1. Supervised learning on available dataset. A black-box tree-based model that required extensive hyperparameters tuning but slightly higher accuracy. �
2. Look-alike modeling uses user similarity in a nutshell. A user is assigned to age group based with probability that depends on a relative distance to the users of each age group. Quick to train, user clusters can be saved locally.

​

7 different models show results 85-94 ROC AUC score. This solution provide posibility build prediction of age and gender for 60 million users.

​

��

Customer

Solution

Results

Inspire brilliant minds to innovate and create.

20

Data consulting / processing automation

1000 employee

x 100 invoices per person

x 2-20 min per document

x 20 days

x 50 parameters

2M invoices per month

0.3M hours per month

100M params per month

​

21

Skills | Knowledge | Collaboration

Salmon’s parasite detection

CHALLENGE:

There are around 4500 salmon farms in Norway. Every week each farm should submit ecology report to the government, otherwise should pay high penalty. Sometimes it’s not possible due to weather and limited access of biologist to the cages.

SOLUTION:

• Underwater capsula with 2 cameras for 24/7 monitoring and for 6k resolution images.
• Image processing algorithm for fish detection and parasites detection on early stages.
• Infrastructure for centralized processing and analytics.

BUSINESS VALUE:

• 10% cost reduction for report generation due to data collection optimisation.
• up to 25% reduction of pesticide use �due to parasite detection on early stages.

​

​

​

​

​

Fruits detection

CHALLENGE:

Client needed to quantify the volume of ripe fruits in the garden using the automatic collection of images. They required the scalable solution showing the applicability of DL-inspired approach.

Accurately locate the fruits in the images with different scales, perspectives, angles etc. and provide robust scalable solution.

SOLUTION:

• Developed object detection algorithm based on Faster R-CNN architecture
• Built AI system that identifies fruit in an image in a nearly human-level performance

BUSINESS VALUE:

• More accurate counting 83% versus 75% (human performance)
• AI framework able to automate counting�on 1.6 million of acres orange fields.

​

​

​

​

Video

Car claim cost evaluation

CHALLENGE:

The Client needed to automatically evaluate �the car repair costs using DL-inspired approach �and eliminate manual evaluation errors.

Each claim contained images with different types �of damages, different perspectives, scale, �amount of dirt, sun glare etc.

SOLUTION:

Developed system classifies damaged parts of a сar, segments the damages in the images and automatically estimates a car claim cost.

BUSINESS VALUE:

Reduce up to 30% costs associated with needs to send� commissar to the car accident with not significant damages

​

​

​

​

​

​

NLP/Chatbots.

25

Skills | Knowledge | Collaboration

Case-XX Customer Support ChatBot

CHALLENGE:

Today’s consumers are active online 24 hours a day, seven days a week. Customers want support when needed, but with around-the-clock online use, that mean high overhead costs in staffing for call centres. More suitable option would be messengers.

SOLUTION:

Developed and released Chatbot solution for Telecom tech support. Chatbot has integrated NLP to understand customers requests and provide related help with multi language support.

​

BUSINESS VALUE:

• Streamline user experience with 24/7 support
• 50% of the call centre enquiries covered by chatbot (~ 600 calls) $2,400 saved/day
• Approx. $864,000/year saved

​

​

​

​

​

Case-4 Virtual assistant in Retail

CHALLENGE:

The Client ask for innovation project to find a way how to leverage from conversation commerce in retail.

SOLUTION:

Smart IoT shell with directed microphone and direct speaker covered a selected zone. Designed smart assistant that convert speech to text and advice customer in his need. Additional camera help to identify products and find similar in a database.

BUSINESS VALUE:

• Increase client satisfaction in pilot shop on 7%
• Enable buying experience for people with disabilities. Blind people could be guided buy sound and their movements controlled with a camera.
• First step in Amazon competition strategy.

​

​

​

​

​

DEEP LEARNING

AS A SERVICE

Skills | Knowledge | Collaboration

28

Skills | Knowledge | Collaboration

Ciklum – Preferred DL Partner for NVIDIA Clients

29

Identification of the need

Data �Collection

Data Preparation

Train�models

Deploy

Improve,�Grow,

Scale

Support

How we deliver deep learning solutions

It’s a unique approach that allows us to help you to identify business needs and apply modern scientific approach to design digital transformation roadmap for you.

​

​

Being NVIDIA preferred deep learning partner we have a library of the best state of the art DNN architectures and frameworks that allow us to start model training immediately as we got data.

Partnership with AWS, Microsoft, IBM and Google Cloud Platform allows us easily setup infrastructure and collect all you data in a right DWH.

​

24/7 support service

Thank You