Slide 1

Slide 1 text

28 Februrary 2018 Deep Learning & Accelerating the NLP Journey in the Unstructured World Jenny Chong, Global Head of eCommunications Surveillance Shahzad Chohan, Global Head of Machine Intelligence and Accelerated Computing

Slide 2

Slide 2 text

eCommunications Surveillance Deep Learning & Accelerating the NLP Journey in the Unstructured World Not for Distribution

Slide 3

Slide 3 text

Our team

Slide 4

Slide 4 text

Team Motto Empower the Art of the Impossible Empower the Art of the Impossible

Slide 5

Slide 5 text

Innovation in Surveillance The Smell of Success Agenda The Surveillance Problem

Slide 6

Slide 6 text

The Surveillance Problem

Slide 7

Slide 7 text

Let’s manipulate the market

Slide 8

Slide 8 text

No content

Slide 9

Slide 9 text

Electronic Communications Surveillance Threat Analysis to mitigate risk Detect potentially malicious and fraudulent activity Detect anomalous behaviour for investigation

Slide 10

Slide 10 text

Forensics Review Resolve Ingest Filter The Surveillance Pipeline

Slide 11

Slide 11 text

Volume of Noise Months Hours Simplicity Chaos Time to Insight Matrix Time Manual Discovery

Slide 12

Slide 12 text

Innovation in Surveillance

Slide 13

Slide 13 text

Surveillance Toolkit Forensics Filter Review Resolve Entity Resolution Artemis

Slide 14

Slide 14 text

Forensics Review Resolve Ingest Filter Artemis

Slide 15

Slide 15 text

Artemis Entity Resolution Artemis Map all communications to the right person accurately Temporal view of a person’s identities across systems Canonical view of a person

Slide 16

Slide 16 text

Entity Resolution Time Volume of Noise Manual Discovery Months Hours Simplicity Chaos Artemis

Slide 17

Slide 17 text

Surveillance Toolkit Forensics Filter Review Resolve Entity Resolution Artemis Natural Language Processing Talos

Slide 18

Slide 18 text

Forensics Review Resolve Ingest Filter Artemis

Slide 19

Slide 19 text

Intelligent Filtering Deep Learning Workflow Parsing Phase Vectorising Phase Algorithmic Phase Stop Words & Stemming Word Embeddings Deep Neural Network “fix market” -> [042 312] “fix car” -> [042 911] “car park” -> [911 020] Machine Learning Workflow Potentially Malicious eComms Classified Malicious eComms “I am fixing the market” “fixing market” “fix market” “fix market” -> “Malign” “fix car” -> “Benign” “car park” -> “Benign”

Slide 20

Slide 20 text

Intelligent Filtering GPU CUDA CUBLAS Python Tensorflow/Theano Keras Magic

Slide 21

Slide 21 text

Entity Resolution Intelligent Filtering Time Volume of Noise Manual Discovery Months Hours Simplicity Chaos Artemis

Slide 22

Slide 22 text

Surveillance Toolkit Forensics Filter Review Resolve Entity Resolution Artemis Natural Language Processing Talos Alert Dashboard

Slide 23

Slide 23 text

Forensics Review Resolve Ingest Filter Artemis Alert Dashboard

Slide 24

Slide 24 text

Entity Resolution Intelligent Filtering Time Volume of Noise Manual Discovery Months Hours Simplicity Chaos Artemis Alert Dashboard

Slide 25

Slide 25 text

Surveillance Toolkit Filter Review Resolve Entity Resolution Artemis Natural Language Processing Talos Alert Dashboard Forensics Cypher Dynamic Visualization of Communication Networks

Slide 26

Slide 26 text

Forensics Review Resolve Ingest Filter Artemis Cypher Alert Dashboard

Slide 27

Slide 27 text

Entity Resolution Intelligent Filtering Time Volume of Noise Manual Discovery Months Hours Simplicity Chaos Artemis Alert Dashboard Cypher Forensics

Slide 28

Slide 28 text

Alert Dashboard Surveillance Toolkit Cypher Dynamic Visualization of Communication Networks Natural Language Processing Talos Review Entity Resolution Artemis

Slide 29

Slide 29 text

Forensics Review Resolve Ingest Filter Artemis Cypher Alert Dashboard

Slide 30

Slide 30 text

Mind the Gap! Producer Consumer Predictive Analytics Surveillance Cost Modelling Catalyst For Success Our Team Our Team

Slide 31

Slide 31 text

The Smell of Success

Slide 32

Slide 32 text

13% Lexicon Search Logistic Regression Machine Learning Naïve Bayes Random Forest Support Vector Machine Target (12.5%) Team Target Deep Learning 37% Number of Alerts

Slide 33

Slide 33 text

Intelligence Engineering Hours Accuracy Heuristic Approaches Generic ML Model Libraries Credit Suisse Model Performance Why such a big difference? • Internal and Domain- Specific Knowledge • Algorithmic and Mathematical Expertise • Reduced time-loss in needless integration activities

Slide 34

Slide 34 text

GPU Innovation 2250 15 Mins per Execution 150 x

Slide 35

Slide 35 text

Deep Learning Toolsets Training | Inference | Emotion | Forensics

Slide 36

Slide 36 text

Over & Out

Slide 37

Slide 37 text

Unicorns are Rare Enterprise is Tough Choose Passengers Carefully Explore without Fear Ecosystem Beats Product Key Learnings

Slide 38

Slide 38 text

Team Motto Empower the Art of the Impossible Empower the Art of the Impossible

Slide 39

Slide 39 text

Deep Learning & Accelerating the NLP Journey in the Unstructured World Q & A Jenny Chong Global Head of eCommunications Surveillance Shahzad Chohan Global Head of Machine Intelligence & Accelerated Computing