Executive interview: Gideon Mann, head of data science, Bloomberg

Cliff Saran Managing Editor 10 Mar 2017 14:15
Data science is one of four focus areas within the CTO’s office at financial news and data company Bloomberg, and the head of data science is Gideon Mann.

Mann professes that his definition of data science is non-traditional. “People define data science in a lot of different ways,” he says. “Bloomberg data science is non-conventional and focuses on three technology areas – natural language processing, information retrieval and search, and core machine learning.”

Arguably, information retrieval and search is the closest fit to conventional data science. Mann says: “Remembering what it was like in the 1990s, you didn’t have Google, Bing or Yahoo and you couldn’t find everything on the internet. Life was quite different.”

But despite the advances in internet search engines, they have their limits, says Mann. “You can often find the document you want, but you may not necessarily find the piece of information you want. If that information lives across many documents, there is nothing that can help you.”

When researching a topic, people used to spend a lot of time going to libraries, reading indices, finding books, compiling – and doing this multiple times. “Now you can assemble all the documents very quickly, but unless it has been preassembled in one place, there is still a step of assessment,” he says.

Mann believes there is far too much information spam and despite the technological breakthroughs in search engines, no one has managed to derive a true sense of understanding the meaning within the mass of information on tap. “The impact of all this contention, on deciding what is true, did something happen or is it an alternative fact means there is a big step from assembling documents to extracting meaning,” he says.

Natural language processing at Bloomberg involves extracting information from written text. For example, Mann says natural language processing could be used to determine the sentiment around a particular company. “There is a lot of information contained in text,” he says. “For example, Kraft makes many, many products. What are they? Traditionally, the approach has been structured data analysis. Now the frontier is looking at extracting this information from multiple libraries.”

Core machine learning

Arguably, machine learning take a very different approach to solving problems compared with programming the computer to calculate the correct results, based on a given dataset. This means that IT people may not naturally turn to machine learning to solve a particular problem, even if it is a good fit for this approach.

But as Mann points out: “If they are talking to me, they have already drunk the Kool Aid.”

Sometimes, says Mann, people’s expectations may be unrealistic, or they are not sure of the constraints of machine learning algorithms. “I don’t think you have to lead people to water. Instead, you have to equip them with the ways of thinking,” he says.

One of the first times Bloomberg used machine learning was for sentiment analysis in 2008-2009. “As a company, we have been doing analytics for a long time,” says Mann. “How to do machine learning is definitely a culture shift.”

Referring to the sentiment analysis project, he said: “They tried conventional approaches and these didn’t work. They were brittle and required a lot of manual effort. But machine learning just worked.”

Since then, the group that looked at machine learning for the sentiment analysis project has now grown and is tackling new areas within Bloomberg, where machine learning is now a core competency. “All of the people we now hire to this kind of work come from this background. It is in the mindset,” says Mann.

“I want to have as many people as possible do machine learning at the company and this requires teaching a lot of different people what it means”Gideon Mann, Bloomberg

Speaking of his role at Bloomberg, Mann says: “I want to have as many people as possible do machine learning at the company and this requires teaching a lot of different people what it means. Some are programmers. We show them the kinds of problem that would work with a machine learning solution, or show them the development cycle and where they may have problems and how to address them.”

He says their applications may experience performance issues, and overcoming these involves collecting more data and retraining the artificial intelligence (AI).

Traditionally, when an application under development is tested and it produces the wrong results, the coders need to go back to the source code and trace where the error occurs. But in machine learning, the wrong result is part of the training, says Mann.

“It is almost like a different set of testing. You have to test for correctness of the algorithm and you have to test for accuracy of the algorithm. And often it is very difficult to disentangle these two things.”

The challenge for the programmer is to determine whether the AI is getting the wrong answer because the algorithm is flawed or whether the data it has ingested for learning is missing some key information, which then affects the AI’s ability to match patterns accurately.

Risky AI

Beyond programming, the big challenges for machine learning include determining whether a project will be successful, says Mann. “You wouldn’t do the project if you had no chance of success and you would probably want a high chance of success, but nothing is ever guaranteed,” he says. “It may be more complicated in ways you don’t expect.”

When starting a new project, it is therefore necessary to begin reducing the risk of failure and assess what is simply unlearnable by machine, says Mann. A machine learning project may or may not work, and project leaders need to be able to address this risk when presenting a business case.

Given that senior business executive may make machine learning a pet project, it is important to manage their expectations, he says. “There are always risks. You come in well prepared, do small experiments, incremental proof of concept, small engagements and small deployments, and gain insights to try to reduce the risks.”

People often approach high-risk projects by building a minimal viable product (MVP), says Mann, and for a machine learning algorithm, the MVP for a machine learning product could be just a facet of the algorithm. “You take it seriously as a risk, then you minimise that risk,” he says.