Solving Real-World Problems with AI #KMWorld

KMWlogo_Stacked_Session Description:

Pittman starts off by discussing the many subfields that make up AI and then looks at how various industries are using it to achieve incredible results. The Raytion speaker shows how machine-learning can be applied for sentiment analysis of unstructured data in the context of social media using an example of a large telecommunications organization. Guarino believes AI is already having a significant impact for the U.S. government (including defense and intelligence community uses cases) and is also providing game-changing capabilities for global enterprises in a range of industries, including financial services, healthcare and all areas of technology. She provides real-world examples of how AI is driving measurable benefits today in a range of industry sectors, discusses the importance of Explainable AI to regulated industries, where being able to justify the reasoning behind algorithmic decisions is essential.

Speakers:Chris Pittman, Principal Security Engineer, Cylance

Christian Puzicha, Senior Solutions Architect, Raytion GmbH

Amy Guarino, COO, Kyndi

Speakers:

[These are my notes from the KMWorld 2018 Conference. I’m publishing them as soon as possible after the end of a session, so they may contain the occasional typographical or grammatical error. Please excuse those. To the extent I’ve made any editorial comments, I’ve shown those in brackets.]

NOTES:

  • Chris Pittman, Chief Engineer, Cylance
    • Data Breaches are expensive.
      • In 2017, the average cost of a data breach is $11.7M. (Health care organizations have the most expensive data breaches.)
      • Ransomware attacks have doubled in 2017 from 13% to 27%. In 2018, we are see 31%-32%.
      • The average time to resolve a maliious insider’s attack is 50 days .
      • The average time to resolve a ransomware attack is 23 days.
    • How does the Security Industry respond to this?
      • Before AI, they worked to understand how the malware and its creator operate, then the security company created a definition file that effectively “innoculates” your network and devices.
      • With the advent of affordable computing, storage, and processing, now we have the technological power to support AI to quickly identify and resolve data breaches. Then, the more we study the behavior of the bad actors, AI can help predict future breaches and help prevent them.  Because of this 99.7% of malware is prevented (accoding to Cylance security).
  • Amy Guarino, COO of Kyndi.
    • Think big, start small, but just get going!
    • If you’re interested in exploring AI, then start with robotic automation (E.g, blueprism) to hep automate repetitive tasks
    • Small data: How do you create a labeled data set? Kyndi ingests unstructured content and then extracts and classifies data embedded in each document. They combine machine learning with natural language process and knowledge graphs to help understand the concept and its context.
    • Ontologies and taxonomies: increasingly, this can be done by machine rather than by hand.
    • Explainability: as the AI system evolves, it helps me make better decisions. However, we have to be able to explain what happened in the “black box” that led to the ultimate decision.
  • Christian Puzicha, Raytion.
  • Self-leaning social media systems
    • Raytion does Enterprise seach (i.e., Google + security)
    • What happens in a social media disaster? (Example, Kendall Jenner’s ad)
    • Raytion did sentiment analysis of the social media response to the Jenner ad
    • Unstructured information (especially on the internet) has special challenges:
      • multiple languages
      • unbalanced input — usually on the unhappy folks complain via social media
      • lots of typos
      • decoding irony and sarcasm
      • the text does not stay stable — if you train the model set, it works now but will not work in 5 years. So you need to keep retraining.
    • How does magical machine learning work?
      • the bad news is that it involves math
      • the good news is that it works — it uses pretty old old math that has been tested over time
      • first, transfer the text into numbers. Then optimize it.
    • Conclusions:
      • different content sources require different modell
      • different languages require different models
      • features of search engines are useful for social media monitoring
      • success depends on size and quality of training data
      • ethics — what to monitor and should I do it?
Share