Machine Teaching for workflow automation RIGA COMM 2020
- 1. Machine Teaching for workflow automation Ariadna Kramkovska Muntis Rudzitis RIGA COMM 2020
- 2. Agenda • Introductions • What is machine teaching? • Benefits, our unique tool, use cases • Change of Paradigm • Demo 1: trained model • Machine teaching quick wins • Automation, document extraction • Demo 2: document classification
- 3. Ariadna Kramkovska Machine Learning Developer at Emergn Muntis Rudzitis Lead Data Scientist at Emergn Presenters
- 4. Emergn’s Machine Learning Lab Our team's core competencies: • Business understanding of how to profit from ML • Machine learning development • Deep learning and reinforcement learning • Data visualization • Algorithms and ML techniques • Data processing, cleaning and preparation OUR TECHNOLOGY STACK We have established and run the largest ML community in the Baltics with 1260 field experts as members. All these activities currently makes our company the No. 1 Choice for young and experience ML talents. Machine learning models Tools and programming languages: • Python • TensorFlow, TensorBoard • R ,OpenCV, Caffe2 • KNIME • Azure Machine Learning Studio • C, C++ Deployment techniques Platforms and environments: • Stand alone models • SAP Hana2 • Microsoft Azure (Cortana intelligences suite) • SQL Server We are partnering with GDEXA to help enable the young generation with highly demanded skills like applied AI/ML, Big Data Analytics and Cloud Applications.
- 5. What is machine teaching? Why use it?
- 6. Business automation challenge From automation with replacement of humans to augmentation and empowerment of subject matter experts. We predict that companies who use augmented automation technologies to empower their environments and educate their people on how to use them for better, more predictable outcomes, will win by providing the best service and building better products. • Wheels for the mind • Find a comfortable level of automation
- 7. Why are we looking for machine learning (ML) alternatives? ML for automation and workflows should: • Be transparent and interactive for business users • Include natural language-based solutions where humans have better comprehension • Understand context and learning from smaller data sets • ML models should be verified and monitored AI director at Facebook YANN LECUN ATARI GAME Self-driving cars need millions of hours of training to reach human level trained in about 20 hours. In 80 hours machine will reach human level aquired by 15 minutes.
- 8. Change of paradigm Key differences • Role of subject matter expert (SME) changes – using our tool, SME trains/provides the logic to improve the model. • SME is integral to the success, needs to be empowered and have the tools to do their work better. • The "one and done" approach is not flexible/doesn't allow for market changes nor incremental knowledge. MACHINE TEACHING MACHINE LEARNING
- 9. Iterative machine teaching process For use cases such as: • Email/text/document classification • Email/text/document anonymization • Entity extraction SME DATA SCIENTIST Initial model training Models could be regularly monitored by SME AUTOMATION WORKFLOW CLASSIFICATION MODEL
- 10. Tools for model quality inspection Machine Teaching administration tools help business users and ML Power Users control classification model quality.
- 11. Demo 1 Monitoring of the trained model and quality check
- 12. Finding quick wins ROUTINE OPERATIONS • RPA Robots • Machine Learning • OCR + Data Extraction REPETITIVE COGNITIVE TASKS COLLABORATION WORKFLOWS • Machine Learning • Rule Engines • Interactive applications • Workflows • Data Enrichment DECISION MAKING AND SUPPORT • AI • Process Mining • Document classification • Automation translation • Collaboration apps • Approvals • Case management • Prioritization of work • Analysis • Data Extraction • Copy data • Enter data • Sort documents FREE UP PEOPLE TIME SUPPORT SHIFT TO DIGITAL OPERATING MODEL TASK AUTOMATION WORKFLOW AUTOMATION DECISION AUTOMATION • Machine Learning Natural Language based use-cases
- 13. Automation of the document flow DOCUMENT LIFE-CYCLE RECEIVE DOCUMENT PROCESS DOCUMENT ARCHIVING • Highly manual • Need decision making, involving knowledge worker to do manual tasks • Text comprehension (SME) • Fraught with errors • Difficult to research/go back, to find things • Time consuming • Not possible when scale is large
- 14. Automation of the document flow DOCUMENT LIFE-CYCLE RECEIVE DOCUMENT SAVE DOCUMENT EXTRA META DATA Document and form recognition using OCR • Be physical document or email or video any format can be input DOCUMENT ROUTING DOCUMENT PUBLISHING ARCHIVING Meta-data extraction • Key fields • Key words • In text • e.g. PO # • Subject • Amount Document classification and routing e.g. • Classification • Invoice routing • Approval, level, direct forward to finance Document anonymization and masking e.g. • Bank and personal info Document archiving by classification e.g. • Archiving based on classification: in cloud or on-site HOT WARM COLD
- 15. Information extraction from documents CLIENT RECORD CLIENT RELATED DOCUMENTS AND KEYWORDS DATA EXTRACTION MODEL *Client information is automatically enriched. Company: BNP Paribas* Title: Name: Tom Barnes* Other information: Meta data Retrieving information from text and documents helps you to obtain data that can and must supplement the information of certain accounting cards. DATA EXTRACTION CONFIG
- 16. Document classification Email • By title • By object • By action Mail • A • B • C Contract Application Invoice CLASS CATALOG The document class is determined by the organizational documents/text classification catalog Training of document class catalogues and document classification algorithm by business user – specialist The classification of documents results in the identification of the document class and can use it to further process the document, usually routing and storing it in an appropriate document process. DOCUMENT CLASSIFICATION MODEL DOCUMENT CLASS
- 17. Document anonymization The purpose of anonymizing documents/text is to cover all or part of sensitive information prior to publication of the document/text. Person 1, social security number, living, Address 1, closed contract Contact Number, with Company ABC for Contract Title DATA MASKING George Bennett, Social security number 989384843*****, address, ***** street * DOCUMENT ANONYMIZATION ANONYMIZATED DOCUMENT ANONYMIZATED TEXT ANONYMIZATION CONFIGURATION ANONYMIZATED DOCUMENT
- 18. Machine Teaching (MT) Tool Machine Learning (ML) Solution Technical architecture SET UP AS AN INTERNAL TOOL ON-PREMISE SERVICE MODULAR STRUCTURE SME/IT: • MT Tool with UI for training & analytics • Custom integrations possible • Endpoints for production scenario provided Data Scientist: • ML model exchangeable - has standardized endpoints • Data and integration – as needed PYTHON CAN BE INTEGRATED INTO EXISTING WORKFLOW MODEL MT DATABASE MODEL DATA ML SOLUTION MT SERVICE CUSTOM INTEGRATION & PRODUCTION MT TOOL
- 19. Machine teaching process PERS LOC ORG DATE PROD LABELING CHANGES VERIFICATION TRAINING/ RETRAINING
- 20. Demo 2 Text classification example
- 21. How to get most value from machine teaching for Natural Language Processing (NLP) and Non-NLP models IN NLP MODELS • Objective – Create Optimal data set for best known optimal models • Use for data labeling and data preparation IN NON-NLP MODELS • Objective – Find optimal/best model with existing data • Use for model monitoring and verification Machine Learning is the subfield of computer science that, according to Arthur Samuel in 1959, gives “computers the ability to learn without being explicitly programmed.” NLP Artificial Intelligence Machine Learning Statistics
- 22. Let's stay connected Contact us at info@emergn.com for an individual demo. Follow us: @emergn MUNTIS RUDZITIS Lead Data Scientist at Emergn https://www.linkedin.com/in/muntis-rudzitis/ ARIADNA KRAMKOVSKA Machine Learning Developer at Emergn https://www.linkedin.com/in/ariadnakramkovska/