NicoElNino - Fotolia
In the near future, some of the biggest gains from AI will come from finding new ways to improve business processes,...
rather than from a blockbuster algorithm.
That's according to Andrew Ng, one of the pioneers behind AI development at Google and Baidu and now CEO and co-founder of Landing AI.
Speaking at the EmTech Digital conference in San Francisco, Ng said he expects much of the economic growth from AI development to happen outside of traditional software industries in fields like retail, transport and logistics. So, Landing AI, founded in 2017, has created an "AI playbook" to help companies in different industries take advantage of AI, a valuable but admittedly difficult technology to implement. A recent Landing AI product is deeplearning.ai, an educational program to help make this AI technique more accessible.
For companies that have not yet developed an AI playbook, Ng recommended they begin with a pilot project -- and be prepared to be met with skepticism from their colleagues. Even at Google, Ng said he had a lot of skeptics when he started implementing deep learning projects. His strategy was to start with a group at Google that was already using other AI techniques (the speech team) to demonstrate the potential for deep learning.
"One of the best things you can do is start a pilot project to deliver quick wins, and that helps teams to develop expertise and gain faith," Ng said.
Here is a summary of his advice on how to get your AI playbook off the ground at non-tech companies.
1. Start by brainstorming
One common mistake that companies make in launching an AI development strategy is to go with the CEO's top pick for a first project. Instead, Ng said a better approach is to start by brainstorming six to 12 ideas and subjecting them to a thorough business and technical review. Business managers should be tasked with doing spreadsheet models of an idea's projected value. The technical due diligence evaluates whether the project is doable in terms of the algorithms and data.
"It is difficult to have good judgments as to what AI can and cannot do," Ng said. CEOs tend to be overly optimistic, in part because the scientific community and the press tend to share more stories about how AI succeeds than where it fails. This leads to a perception that AI can do anything.
Ng believes the C-suite perception of AI development has led to unrealistic expectations around, for example, chatbots that have resulted in a lot of failed projects. In order to help reduce these AI myths, Ng worked on a free course program called AI for Everyone to make basic AI concepts more understandable to nontechnical people. One aspect of the course involves identifying examples of what AI can and cannot do.
"I don't think every company can deliver huge wins in the short term, but many can," Ng said.
The program also includes guidance on how to think about the kind of data required to make AI work well. Just because a company has a data lake, it does not mean the data is suitable for AI. It is not true that having more data will always guarantee winning, Ng said. One healthcare company invested a lot of money acquiring healthcare companies for their data. But Ng said they have not found a way to use this data in a meaningful way.
2. Realize that AI breaks in new ways
Andrew NgCEO, Landing AI
Another key to getting an AI playbook off the ground is understanding how AI systems can fail. One of the challenges of machine learning is that the workflows are different. As a result, AI researchers are still trying to figure out how to communicate to businesses the system conditions that can lead to a failure, which, in some cases, can have dire consequences: a failed medical diagnosis or a crashed airplane.
For example, the press is quick to pick up on how a new AI algorithm can deliver more reliable X-ray evaluations than humans. But Ng pointed out that these reports typically neglect mentioning that the AI only works on the latest X-ray equipment with perfectly captured images. Doctors outperform AI on older X-ray machines, or when the patient is imaged at an angle different from the one the AI was trained on.
Thus, a new AI system ends up not working in the real world because businesses have not created a good process for addressing these "boundary" conditions, Ng said. Traditional software engineering has developed processes like Agile, scrum and code review to address these problems. But the AI world is still trying to figure out how to best implement comparable processes.
3. Get ready for small data
It's important to be thoughtful about how you acquire your data. Although large data sets can be useful for certain applications, big data for its own sake is slightly overhyped, Ng said. In the short run, many enterprises could see greater benefit from AI development by building systems that work with small data.
Most of the AI work has focused on building algorithms that work with larger data sets, because it is relatively easy. But it is much more challenging to build algorithms that can be trained on a few hundred images than a few million images. For example, Ng's team is working on a project to use machine vision to analyze manufacturing defects in cell phones, and (fortunately) these companies have not manufactured millions of broken phones to analyze.
So, they are working on developing algorithms for detecting bad products from a few hundred images. Other projects are looking at analyzing pictures of weeds in a field and many healthcare images of rare problems.
"That capability is on the bleeding edge of where AI is going," Ng said. Companies that develop the capabilities that launch projects on small data sets will be in a better position than competitors that rely on larger data sets.
4. Build AI partnerships, understanding across the business
The shortage of AI talent is a problem faced by many businesses. Ng said that his secret plan behind writing about the fundamentals of AI guide is to actually make the AI talent crisis more acute: Once people across the business have a clear understanding of how AI development can provide value in a particular business unit, he said they will want to hire more engineers. "I have told engineers that if you want your company to adopt AI, one of the best things you can do is get your C-suite to take AI for Everyone," he said.
5. Measure the impact of AI
Ng believes that online advertising tracking is not the most inspiring application of AI, yet the impact of AI is easily measured in this area. The value of AI can be harder to measure for other AI use cases. Still, Ng advocated that companies consider many metrics such as reducing cost by using AI, or generating value with new AI business offerings, or measuring revenue trends in business units that adopt AI -- but be prepared for the difficulty of tracking these critical business metrics. For example, if AI improves the user experience, it can be hard to connect this improvement to revenue, because changes in revenue can be due to many factors.
6. Keep it simple
Ng is bearish on the prospects of cutting-edge AI research into domains like quantum computing and artificial general intelligence, which, by the way, he predicted could be anywhere from 500 to 500,000 years away. In the meantime, he sees enough opportunities where AI could increase business impact in one to two years, and he has decided to focus on these. "The only thing better than a huge long-term [AI] opportunity is a huge short-term opportunity, and we have a lot of those now," he said.
He expects to see some of the biggest economic gains outside of Silicon Valley by existing industries making their business processes a bit more efficient, adding that a coffee company, for example, is likely to see more gains by developing an AI strategy than by using search engine optimization.
In the short run, Ng suggested that every company find a way to get started with a small project with a couple of engineers, rather than a big 10-year plan. "That teaches your organization what it feels like to do these AI projects and gain momentum," he said.