From our attitude towards work to our grasp of what two metres look like, the coronavirus pandemic has made us rethink how we see the world. But while we’ve found it hard to adjust to the new reality, it’s been even harder for the narrowly-designed artificial intelligence models that have been created to help organisation make decisions. Based on data that described the world before the crisis, these won’t be making correct predictions anymore, pointing to a fundamental problem in they way AI is being designed.
David Cox, IBM director of the MIT-IBM Watson AI Lab, explains that faulty AI is particularly problematic in the case of so-called black box predictive models: those algorithms which work in ways that are not visible, or understandable, to the user. “It’s very dangerous,” Cox says, “if you don’t understand what’s going on internally within a model in which you shovel data on one end to get a result on the other end. The model is supposed to embody the structure of the world, but there is no guarantee that it will keep working if the world changes.”
The COVID-19 crisis, according to Cox, has only once more highlighted what AI experts have argued for decades: that algorithms should be more explainable.
SEE: How to implement AI and machine learning (ZDNet special report) | Download the report as a PDF (TechRepublic)
For example, if you were building a computer program that was a complete blackbox, aimed at predicting what the stock market would be like based on past data, there is no guarantee it’s going to continue to produce good predictions in the current coronavirus crisis, he argues.
What you actually need to do is build a broader model of the economy that acknowledges supply and demand, understands supply-chains, and incorporates that knowledge, which is closer to something that an economist would do. Then you can reason about the situation more transparently, he says.
“Part of the reason why those models are hard to trust with narrow AIs is because they don’t have that structure. If they did it would be much easier for a model to provide an explanation for why they are making decisions. These models are experiencing challenges now. COVID-19 has just made it very clear why that structure is important,” he warns.
It’s important not only because the technology would perform better and gain in reliability, but also because businesses would be far less reluctant to adopt AI if they trusted the tool more. Cox pulls out his own statistics on the matter: while 95% of companies believe that AI is key to their competitive advantage, only 5% say they’ve extensively implemented the technology.
While the numbers differ from survey to survey, the conclusion has been the same for some time now: there remains a significant gap between the promise of AI and its reality for businesses. And part of the reason that industry is struggling to deploy the technology boils down to a lack of understanding of AI. If you build a great algorithm but can’t explain how it works, you can’t expect workers to incorporate the new tool in their business flow. “If people don’t understand or trust those tools, it’s going to be a lost cause,” says Cox.
Explaining AI is one of the main focuses of Cox’s work. The MIT-IBM Watson Lab, which he co-directs, comprises of 100 AI scientists across the US university and IBM Research, and is now in its third year of operation. The Lab’s motto, which comes up first thing on its website, is self-explanatory: “AI science for real-world impact”.
Back in 2017, IBM announced a $240 million investment over ten years to support research by the firm’s own researchers, as well as MIT’s, in the newly-founded Watson AI Lab. From the start, the collaboration’s goal has had a strong industry focus, with an idea to unlock the potential of AI for “business and society”. The lab’s focus is not on “narrow AI”, which is the technology in its limited format that most organizations know today; instead the researchers should be striving for “broad AI”. Broad AI can learn efficiently and flexibly, across multiple tasks and data streams, and ultimately has huge potential for businesses. “Broad AI is next,” is the Lab’s promise.
The only way to achieve broad AI, explains Cox, is to bridge between research and industry. The reason that AI, like many innovations, remains stubbornly stuck in the lab, is because the academics behind the technology struggle to identify and respond to the real-world needs of businesses. Incentives are misaligned; the result is that organizations see the potential of the tool, but struggle to use it. AI exists and it is effective, but is still not designed for business.
SEE: Developers say Google’s Go is ‘most sought after’ programming language of 2020
Before he joined IBM, Cox spent ten years as a professor in Harvard University. “Coming from academia and now working for IBM, my perspective on what’s important has completely changed,” says the researcher. “It has given me a much clearer picture of what’s missing.”
The partnership between IBM and MIT is a big shift from the traditional way that academia functions. “I’d rather be there in the trenches, developing those technologies directly with the academics, so that we can immediately take it back home and integrate it into our products,” says Cox. “It dramatically accelerates the process of getting innovation into businesses.”
IBM has now expanded the collaboration to some of its customers through a member program, which means that researchers in the Lab benefit from the input of players from different industries. From Samsung Electronics to Boston Scientific through banking company Wells Fargo, companies in various fields and locations can explain their needs and the challenges they encounter to the academics working in the AI Watson Lab. In turn, the members can take the intellectual property generated in the Lab and run with it even before it becomes an IBM product.
Cox is adamant, however, that the MIT-IBM Watson AI Lab was also built with blue-sky research compatibility in mind. The researchers in the lab are working on fundamental, cross-industry problems that need to be solved in order to make AI more applicable. “Our job isn’t to solve customer problems,” says Cox. “That’s not the right use for the tool that is MIT. There are brilliant people in MIT that can have a hugely disruptive impact with their ideas, and we want to use that to resolve questions like: why is it that AI is so hard to use or impact in business?”
Explainability of AI is only one area of focus. But there is also AutoAI, for example, which consists of using AI to build AI models, and would let business leaders engage with the technology without having to hire expensive, highly-skilled engineers and software developers. Then, there is also the issue of data labeling: according to Cox, up to 90% of the data science project consists of meticulously collecting, labeling and curating the data. “Only 10% of the effort is the fancy machine-learning stuff,” he says. “That’s insane. It’s a huge inhibitor to people using AI, let alone to benefiting from it.”
SEE: AI and the coronavirus fight: How artificial intelligence is taking on COVID-19
Doing more with less data, in fact, was one of the key features of the Lab’s latest research project, dubbed Clevrer, in which an algorithm can recognize objects and reason about their behaviors in physical events from videos. This model is a neuro-symbolic one, meaning that the AI can learn unsupervised, by looking at content and pairing it with questions and answers; ultimately, it requires far less training data and manual annotation.
All of these issues have been encountered one way or another not only by IBM, but by the companies that signed up to the Lab’s member program. “Those problems just appear again and again,” says Cox – and that’s whether you are operating in electronics or med-tech or banking. Hearing similar feedback from all areas of business only emboldened the Lab’s researchers to double down on the problems that mattered.
The Lab has about 50 projects running at any given time, carefully selected every year by both MIT and IBM on the basis that they should be both intellectually interesting, and effectively tackling the problem of broad AI. Cox maintains that within this portfolio, some ideas are very ambitious and can even border blue-sky research; they are balanced, on the other hand, with other projects that are more likely to provide near-term value.
Although more prosaic than the idea of preserving purely blue-sky research, putting industry and academia in the same boat might indeed be the most pragmatic solution in accelerating the adoption of innovation and making sure AI delivers on its promise.