Today marked the conclusion of VentureBeat’s Transform 2020
summit, which took place online for the first time in our history. (Here’s a link to the livestream
on YouTube, which will wrap up later this afternoon.) Luminaries including Google Brain ethicist Timnit Gebru and IBM AI ethics leader Francesca Rossi spoke about how women are advancing AI and leading the trend of AI fairness, ethics, and human-centered AI. Twitter CTO Parag Agrawal detailed the social network’s efforts to apply AI to detect fake or hateful tweets. Pinterest SVP of technology Jeremy King walked through learnings from Pinterest’s explorations of computer vision to create “inspirational” experiences. And Unity principal machine learning engineer Cesar Romero brought clarity to the link between synthetic data sets and real-world AI model training.
That’s just a sampling of the panels, interviews, and discussions to which Transform 2020 attendees had front-row seats this week. But the sessions that caught my eye were those touching on practical, tangible AI applications as opposed to theoretical. Research remains crucial to the field’s advancement, and there’s no sign it’s slowing – the over 1,000 papers accepted to ICML 2020 suggest the contrary. However, production environments are perhaps the best opportunity to battle-test proposed tools and algorithms for robustness. Outcome predictions are just that: predictions. It takes real-world experimentation to know whether hypotheses will truly pan out.
Barak Turovsky, Google AI director of product for the natural language understanding team, elucidated steps
Google took to mitigate gender bias from the language models powering Google Translate. Leveraging three AI models to detect gender-neutral queries and generate gender-specific translations before checking for accuracy, Google’s system can provide multiple responses to translations of words like “nurse” and let users choose the best one (e.g., the masculine “enfermero” or the feminine “enfermera”). “Google is a leader in artificial intelligence, and with leadership comes the responsibility to address a machine learning bias that has multiple examples of results about race and sex and gender across many areas, including conversational AI,” Turovsky said.
Like Google, software company Cloudera doubled down
on productization of its AI and ML technologies. Senior director of engineering Adam Warrington said it deployed a chatbot to improve customer question-and-answer experiences in under a month, leveraging proprietary data sets of client interactions, community posts, subject-matter expert guidance, and more. The underpinning models can understand relevant words and sentences within a support case and extract the right solution from the best source, whether a knowledge base article, product documentation, or community post.
For Yelp, deployment is a core part of the experimentation process, enabled by the company’s Bunsen platform. Using Bunsen through a frontend user interface called Beaker, data scientists, engineers, execs, and even public relations reps can determine whether products and models have any negative impact on the growth of business metrics or if they’re meeting goals. Yelp employees get the scale of being able to deploy a model to a cohort of users depending on how they want to reach them, as well as the flexibility to determine if the functionality is perhaps not optimal or, worst-case scenario, is harmful. “We have a rapid way of turning those experiences off and doing what we need to do to fix them on the backend,” Yelp head of data science Justin Norman told VentureBeat. “One of the best things about what Bunsen allows us to do is to scale at speed.”
When it comes to practical uses of AI and machine learning in the financial sector, Visa is at the forefront
with projects that demonstrate the potential of these technologies. As a rule, the company looks for use cases where AI and ML could deliver at least a 20% to 30% efficiency increase. Its Visa Advanced Authorization platform is a case in point: It uses recurrent neural networks along with gradient boosted trees to determine the likelihood transactions are fraudulent. Melissa McSherry, a senior vice president and global head of data for Visa, said the company prevents $25 billion in annual fraud thanks to the AI it developed. “We have definitely taken a use case approach to AI,” she said. "We don’t deploy AI for the sake of AI. We deploy it because it’s the most effective way to solve a problem.”
AI has a role to play
in health care, as well. CommonSpirit Health, the largest not-for-profit health care provider in the country, is applying models to optimize the rounds its doctors and nurses make every day. “Based upon our thousands of analysis of patients, [if] we don’t address the patient in room seven first, they’re going to have to stay longer than they would need to otherwise,” chief strategic innovation officer Rich Roth explained. “Using AI that way, really to accelerate our workflow, and to clearly show to our caregivers the clinical benefit of why that data is important, is a great example of how technology can help enhance care.”
Thanks for reading,
AI Staff Writer