Knowledge Representation and Reasoning into Machine and Deep

Knowledge Representation and Reasoning into Machine and Deep Learning David Newman SVP, Innovation R&D Innovation Group, Wells Fargo Bank david. newman@wellsfargo. com linkedin. com/in/davidsnewman 1 September 27, 2017

Presentation Topics The Challenge of Data What is Knowledge Representation and Reasoning (KRR)? What are some ways that KRR via Ontologies can be Applied to Improve Machine and Deep Learning? What are Some Use Cases? 2

Collecting, Cleaning, and Organizing Data Consumes at Least 80% of Data Scientists Time. Why are we so Impacted by Data Challenges? *Cleaning Big Data: Most Time-Consuming, Least Enjoyable Data Science Task, Survey Says, Forbes, March 2016 3

What are Some of these Key Data Challenges? Data challenges include: ü reconciling and harmonizing disparate data across line of business silos ü linking and aggregating data ü validating, curating, and classifying data for operational processing and creating good feature sets for machine learning We need an effective data management capability that can ensure that data: ü is harmonized and integrated across all data sources ü is curated and aligned to common meaning ü is utilized and understood not only by humans, but by machines as well 4

Conventional Data Management Capabilities are Not Sufficiently Fulfilling our Data Needs! Bad Data 5

Applying Knowledge Representation and Reasoning using Semantic Technology is a Way to Meet our Complex Data Needs! 6

Ontologies Express the Meaning of Concepts Concept of “Corporate Control” Subject Corporation Predicate controls Object Corporation is kind of type controls majority voting shares plays role Parent Company inf is inverse of ere n ce ce ren e f n London Bank plays role i is majority controlled by “Things” not Strings Subsidiary 7

For Finance the EDM Council is Developing a Free and Open Source Common Financial Data Standard Using Ontologies 8

FIBO Ontologies Provide a Scaffolding for Concept Reuse and Extension… Supporting Efficiency, Effectiveness and Governance Industry Standard Ontology Enterprise Knowledge Graph 9

Ontologies can Play a Critical Role to Better Enable AI Artificial Intelligence Knowledge Representation and Reasoning Machine Learning Deep Learning Natural Language Processing 10

Ontologies are a Bridge Between Human Induced Knowledge and Machine Induced Knowledge k c a g g n i s s re o r P Up e h t A t S I Cognitive Computing Machine and Deep Learning Enterprise Ontology Industry Ontology Human Induced Knowledge ü Foundational knowledge for cognitive computing to learn and build stronger associative connections ü Semantically curated and inferred data will give greater lift to machine learning ü Prior knowledge from ontologies used for supervised ML training sets for data quality, correctness and better NLP ü Conceptual scaffolding for enterprise ontology ü Validate, curate, link and harmonize legacy data 11

Ontologies can be Used for Better Feature Engineering Disparate Data Sources Data Linkage Data Validation Data Classification and Harmonization and Consistency and Inference Create Feature Set Identify instances Smart Dimensionality From Multiple of incorrect or Reduction Disparate Data inconsistent data Sources and Formats 12

Ontologies Enable Data Harmonization and Alignment Needed for Consolidation of Disparate Data Based on Common Meaning Global Bank owns > 50% voting shares of London Bank Semantic Mapping RDBMS Semantic Adapters FIBO Operational Ontologies Big Data 13

Ontologies Better Position us to Manage Linked Data and Reduce Data Redundancy “Things Not Strings” 14

Using Ontologies … Automated Logical Consistency Checks can be Performed to Identify Violations of Semantic Rules in Features Definition of Aunt in Ontology Supervised Machine Learning Data in Message Payload Reasoning Logical Inconsistency! 15

Ontologies can be Used to Reason over Data for Classification and Dimensionality Reduction Fixed Float IR Swap Business Entity Human Facing Definition An interest rate swap in which fixed interest payments on the notional are exchanged for floating interest payments. Interest Rate Swap London Bank type Inferred identifies Fixed Float IR Swap type LEI 5001 Swap Leg 1 has principle 10000000 has currency USD Swap has rate LIBOR eg h Interest. Rate. Swap and has. Leg some Fixed. Rate. Leg and has. Leg some Floating. Rate. Leg identifies LEI 7777 s. L Inferred type eg Swap 1001 L as Machine Facing Definition Atlas Bank Inferred ha Floating Rate Leg s ha rty pa LEI Fixed Float IR Swap Business Entity Swap Leg 2 has rate 5% s ha rty pa LEI Inferred type Fixed Rate Leg has principle 10000000 has currency USD 16

Knowledge Graph is the Convergence of Ontologies with Machine Learning Insights Probabilistic Associations and Classifications Knowledge Graph all counterparties for interest rate swap trades that are likely (> 99. 5%) to default given a x% rise in LIBOR Operational Semantic Graph Data 17

Machine Learning can Generate Probabilistic Associations that can be Expressed by the Knowledge Graph as New Relations Address Explanation Other Person A Name Cosine Customer A Similarity Household Other Relation Infer predicate relationships (same as, colludes, knows, et. al. ) Customer B Address Person B Name From Strings to Things Enterprise Entity Resolution Dedupe/Link Customers, People, Companies, Addresses KYC, Fraud, Credit Risk … Other Relation 18

Ontologies can be Used to Transform Unstructured Content into a Structured Form for Machine Learning and in silico Memory “I am planning to buy a truck to expand my construction business” capture content James Jackson ically semant ify text s s a l c d n parse a Natural Language Understanding Intent: purchase truck Context: business expansion Action: recommend Auto Loan Action: offer to increase credit line plans to expand intended use stor e com custo m mu nica er tion s understanding recommendation machine learning Knowledge Graph 19

Ontologies can also Help Generate More Precise Neural Word Embeddings to Better Optimize Natural Language Understanding Definition Concept Web Link Ontology Concept(s) Relation Definition Concept Annotation Web Link ü Understanding contracts that describe financial instruments ü Merge human knowledge Ontology Concept Iteration for Text Generation and Extraction with machine knowledge ü Train with less but more precise data Generate Neural Word Embeddings for each Concept in Ontology Neural Network Neural Word Embeddings 20

Important Advances in Ontology and Machine Learning Research from the University of Mannheim, Germany [Paulheim, Stuckenschmidt] show ontologies can help ML perform validation of data at least 50 x faster than a state of the art semantic reasoner by training a binary classifier. Research from Rensselaer Polytechnic Institute [Makni, Hendler] shows how we can use ontologies to train deep learning neural networks to perform reasoning to materialize semantic inferences efficiently at scale. Research from the University of Oregon [Wang, Dou, Lowd] shows how we can use ontologies to improve ML insights by building a lattice of deep neural networks that reflects the structure of taxonomies. They call this Semantic Deep Learning. 21

Thank You! david. newman@wellsfargo. com linkedin. com/in/davidsnewman 1 22