We don’t have data that represents the new normal
Artificial intelligence can’t keep up with the coronavirus.
Many companies are therefore having to rethink the complex models they use for the all-important job of forecasting their sales and budgets. The data and assumptions baked into those statistical or machine-learning models are now out of date because of the recent economic upheaval, making the conclusion they provide suspect.
“Our historical, analytical data—what we used in the past—is out the door,” Paul Cormier, CEO of IBM-owned business software maker Red Hat, recently told Fortune. “I think everyone feels that way.”
Doug Merritt, the CEO of IT firm Splunk, shares Cormier’s sentiment. Improving the accuracy of corporate forecasting is something “we are all struggling with,” he said. Although Merritt said models provide accurate short-term forecasts, it’s challenging to develop ones that look 30 to 120 days down the road.
As Rob Thomas, the senior vice president of IBM’s cloud and data platform, said, “In any instance, your models are only as good as the data.”
“When something happened that didn’t happen before, most organizations don’t have data that represents the new normal,” Thomas said. On the flip side, he explained that as companies continue collecting more data during this pandemic, they’ll have a better understanding about this “new normal.”
Thomas believes that companies should “never throw out data” and “never throw out models.” They still carry “institutional” knowledge, and if a company were to build newer models from scratch, they would lose some historical insights built into the older formulas, he explained.
If anything, the coronavirus pandemic has meant that now, more than ever, there needs to be an actual person scrutinizing financial models and not just technology, Thomas explained. Savvy data scientists and business analysts need to use their “human intuition” and reasoning skills to assess how much worth should be placed on the predictions of the impacted models.
“You see a lot of people doing that.” Thomas said.
As for whether companies should expect machine learning or other predictive systems to automatically adjust to sudden upheavals like the coronavirus, it’s unlikely to happen any time soon.
“That would be the Holy Grail,” Thomas said.
P.S. We’d love to hear any interesting methods you or your company may be implementing to make predictive models more robust during this coronavirus pandemic. Feel free to email me—I’m interested in hearing your story. (Please no vendor pitches!)
A.I. IN THE NEWS
Bye Bye fingerprints, hello facial recognition. Biometric technologies related to the “fingerprint and vein” have been “dealt a substantial blow” because of government policies intended to mitigate physical contact amid the coronavirus pandemic, according to the advisory firm ABI Research. The research firm also said that the biometrics industry is excited about improvements in “face and iris recognition” that will allow for better “authentication, identification, and surveillance operations for users and citizens wearing protective headgear, face masks, or, with partially covered faces.” Chinese tech companies SenseTime, Megvii, Alibaba, and Baidu are leading investments in this area, ABI said.
Mercedes speeds towards A.I. Mercedes-Benz is using natural language processing techniques to improve the company’s MBUX infotainment system for its automobiles, CNET reported. Mercedes is deciding on a possible option for users that would remove its wake-word command “Hey, Mercedes” to activate the service, which the article states “means the MBUX assistant will always be listening, chiming in when the AI algorithm deems it appropriate.” A Mercedes executive said the company “will be very careful on that one" and “will offer something only if the customer wants it.”
IBM pushes into “AIOps.” IBM debuted new services related to a trendy kind of IT monitoring technology that analysts have referred to as AIOps. The term acts as a catch-all referring to using machine-learning technology to analyze server data like “log data” in the hopes of spotting errors and preventing crashes before they occur.
Investors are digging robotic software. Covariant, a startup developing robotic gripping software, got $40 million in funding from investors including Index Ventures, Radical Ventures, and Amplify Ventures. Covariant recently partnered with industrial robotics giant ABB to develop more capable robots, targeting the logistics space in particular.
EYE ON A.I. TALENT
Twitter added deep-learning expert Fei-Fei Li to the company's board of directors. Li, a co-director of Stanford University's Human-Centered AI Institute, oversaw the milestone ImageNet computer-vision contest that helped spur a renaissance in deep learning. She was also previously the chief scientist of A.I. and machine learning at Google, but left shortly after news organizations started reporting on Google's internal struggles over potential A.I. projects with the Pentagon.
Najat Khan said via LinkedIn that she is now the chief data science officer at the research and development arm of Janssen, a pharmaceutical firm owned by Johnson & Johnson. Khan was previously the chief operating officer of Janssen’s R&D data sciences unit and was a senior principal at Boston Consulting Group.
Giancarlo Miluccio is leaving L'Oréal as the personal care company’s chief data officer based in France. Miluccio announced his departure in a LinkedIn post but did not say where he is going.
EYE ON A.I. RESEARCH
Reinforcement learning meets economics. Researchers from Salesforce and Harvard University published a non-peer-reviewed research paper detailing how reinforcement learning—when software learns through repetition—could be applied to the field of economics.
Salesforce chief scientist Richard Socher told Fortune that the paper is the “most impactful project I’ve ever worked on.” The paper describes how Salesforce created a deep-reinforcement-learning system that was able to devise an equitable tax policy within the confines of a custom-built video game. The A.I. system’s tax policy performed better than other tax policies including the free-market system, the U.S. federal single-filer 2018 tax schedule, and the Saez tax framework, according to the paper’s authors.
Salesforce researchers were particularly pleased that their A.I. system was able to take in account factors including the notion that people may not feel like working too much if they don’t have to pay taxes. It also takes in account that while people may enjoy working for money, there’s a limit to how much they’re willing to put in on the job—after all, people do like their weekends.
Socher concedes that the research is still in its infancy, and it’s not realistic to assume real-life economists are going to be incorporating the technology into their practice. For one, the reinforcement-learning system was built to learn from a relatively primitive video game, underscoring how current RL systems are limited by the digital environments they’re trained in. There are only four “agents” in the video game, which are not representative of the billions of people (or millions of companies) around the world who have to pay taxes each year. And the game’s economy is simplistic, consisting of the four agents doing tasks like collecting stones and wood, trading those resources, and building homes.
Additionally, the A.I. system doesn’t take in account the political factors that are often the reason why lawmakers choose to implement certain tax policies. I’d personally love to see A.I. agents that represent companies as well as virtual recreations of tax havens like the Cayman Islands.
Future research will include building a more complex video game and increasing the number of A.I. agents, Salesforce researchers said. One benefit of the company’s virtual economy is that the A.I. system can learn from “millions of years of simulation” that wouldn’t be possible in the real world, Socher said. Still, it takes a tremendous amount of computing power to run those millions of years of simulation, and Salesforce researchers declined to comment on how much it cost to train their A.I. system.
Ultimately, Socher said he’s hopeful that the A.I. system will improve over time and act as a starting point for economists interested in experimenting with reinforcement-learning techniques. The research marks another example of researchers applying reinforcement-learning techniques to areas besides competitive video games where one player needs to beat another.
FORTUNE ON A.I.
How A.I. may help solve science’s ‘reproducibility’ crisis—By Jonathan Vanian
Alphabet vets raise $400 million to remake America’s infrastructure—By Jeff John Roberts
In defense of Elon Musk—By Aaron Pressman
The diverging paths of two self-driving car pioneers. Self-driving car pioneers Bryan Salesky and Anthony Levandowski are the subjects of two different profiles by Wired and Automotive News, respectively. Salesky, the CEO of Argo AI, has essentially become Ford’s golden child as the auto giant pursues self-driving vehicles. In 2017, Ford pledged to invest $1 billion in Argo AI over a five-year period. Meanwhile, Levandowski has become “the personification of Silicon Valley’s more unseemly impulses,” according to Automotive News, as he recently plead guilty of taking corporate documents from Google (where he once worked) to Uber.
As Automotive News puts it: “Recently divorced, bankrupt and an admitted thief, Levandowski has seen much of the legacy he created now in tatters.”
Contrast Levandowski’s epic fall with the following critique of Salesky (also a former Google employee) in the Wired piece: “Some former Google colleagues charge him with a lack of charisma, which they deem a handicap when it comes to motivating a team to overcome a brutally difficult obstacle, even if they don’t have to worry too much about 2021.”
Salesky’s greatest fault is that he’s boring, which, compared to Levandowski’s antics, doesn’t sound too shabby.
Update May 12: Story corrected to say Ford's investment was $1 billion over a five-year period.