The ARF’s Paul Donato and Henry Wolf closed with a summary of best practices, as outlined in a recent ARF Knowledge at Hand:
- Make sure you have clearly defined the problem that AI should solve
- Identify the right talent for the situation
- Ensure that you use data in large enough quantities to solve the problem (This may be the most important practice)
- Use algorithms suitable to solve the problem
- Capitalize on open source resources: you don’t have to start from scratch
- Adhere to an AI code of ethics, an increasingly important part of AI. They stated that the ARF will soon release a code of conduct that will include AI-related principles
- Choose AI solution providers carefully – and talk to several; they differ
Monica Fogg – Offering Manager Watson Ads, IBM Watson Ads.
She believes that AI has to satisfy four features. It must be able to:
- Understand the world the way in which humans understand the world, i.e., be able to deal with unstructured data, not have to rely on specific inputs such as those in a spreadsheet
- Reason, in other words make logical sense out of what it understands
- Learn from experiences and improve over time
- Interact, which meant to be able to communicate with humans by producing words, voice, and/or images that resonate
Ben Royce, Head of Performance Data Science, Global Agency at Google
He showed how Google is applying AI in advertising, using two examples.
- Google is exploring AI capabilities to distinguish emotions and then applying the learning to see if AI can understand concept of a brand. An early finding: the difficulty of identifying an emotion via AI varies, as it does with humans, e.g., happy and sad are easier sentiments to recognize than bittersweet and sarcasm.
- To test how AI might associate emotion with brands, Google sought to learn how AI would group different brands without relying on logo recognition. They started by taking the content from 100 brands’ websites to see if AI could associate different brands in a category. They found: AI was successful when the brands communicated similar points, e.g., it could group different telecommunications companies, because they all talked plan features. Other times it produced unexpected groupings, perhaps because it has no context.
Source: The ARF Leadership Lab: AI (2018, September 8). The ARF.