Presented By O’Reilly and Intel AI
Put AI to work
Sep 4-5, 2018: Training
Sep 5-7, 2018: Tutorials & Conference
San Francisco, CA

Productionalizing deep learning for computer vision

Labhesh Patel (Jumio)
4:00pm-4:40pm Friday, September 7, 2018
Implementing AI, Interacting with AI
Location: Continental 1-3
Secondary topics:  Computer Vision, Platforms and infrastructure

Who is this presentation for?

  • Data scientists, AI specialists, and C-level management

Prerequisite knowledge

  • A basic knowledge of deep learning and its potential applications in computer vision and the basic applications of machine learning workflows

What you'll learn

  • Discover how Jumio is productionalizing DL to better extract data from ID documents, better detect fraud, and better inform our risk scoring engine


Jumio has processed more than 100 million IDs over the last few years. The company uses this massive dataset to recognize important patterns—patterns often indistinguishable by the human eye. Labhesh Patel explains how deep learning is informing Jumio’s computer vision through smarter data extraction, fraud detection, and risk scoring.

Topics include:

Data extraction: Jumio has created JADE (Jumio Automated Data Extraction), a tool that leverages deep learning to “dewarp” or rectify an image, exploit multiline OCR for reading lines of text, and field finding (e.g., where we should expect to find a person’s first name on a UK driver’s license). Because Jumio processes hundreds of thousands of IDs every day, it can feed its deep learning algorithms with lots of data to improve their ability to recognize specific ID documents and know how to extract the data and make sense of it.

Fraud detection: Deep learning helps identify potentially fraudulent IDs in several important ways. DL helps determine if an ID has been manipulated in any way (e.g., if a digit has been added). For instance, since Jumio has seen hundreds of thousands of Nepalese driver’s licenses, it knows which fonts are used, where the picture should be placed, how many digits are in the DL number, where the security features are placed, etc. If anything has been manipulated or changed, it is flagged for closer review.

Risk scoring: Because of the volume of global verifications Jumio has performed and the fact that every verification is confirmed or denied by a human agent (the company’s verification experts), it can leverage deep learning to identify patterns and probability of fraud based on a combination of high-risk variables. Without giving away too much of its secret sauce, Jumio knows through its deep learning algorithms the countries, the ID types, the methods of ID capture, and a handful of other variables that result in the highest incidences of fraud. In fact, the company knows from its analysis that 1% of IDs account for about 15% of known fraud. Jumio uses this scoring to alert its verification experts to pay special attention to these high-risk IDs.

Photo of Labhesh Patel

Labhesh Patel


Labhesh Patel is CTO and chief scientist at Jumio, where he’s responsible for driving the company’s innovation in the identity verification space with deep learning, computer vision, and augmented intelligence—an alternative conceptualization of artificial intelligence that focuses on AI’s assisted role to enhance human intelligence. An accomplished leader with over 15 years of experience in corporate and entrepreneurial settings, Labhesh has proven experience leading engineering teams, launching new online services (from concept creation to customer delivery), and developing ground-breaking technologies at companies including Cisco, Abzooba,, Spotsetter, and CellKnight. He has 175 patents filed with another 134 patents issued under his name. Labhesh holds an MS in electrical engineering (MSEE) from Stanford University and a BT from the Indian Institute of Technology in Kanpur.