According to Deloitte Global's 17th edition of the Technology, Media & Telecommunications (TMT) Predictions research, global organizations will double their use of machine learning technology by the end of 2018, with High Tech, Communications, and Financial Services set to lead the charge. However, while many are indeed looking to adopt machine learning, the current reality appears to be very different. In one of many surveys to find slow rates of adoption, Belatrix Software found that just 18% of companies asked if they had already started a machine learning initiative had done so, 40% that they were investigating it but hadn’t started, and 42% that they had no plans to start one at all.
The reality is that there are still a number of challenges that have to be overcome before the technology has a real impact. We asked seven leading experts in the sector from some of the world's leading organizations what they believed were the biggest obstacles machine learning has to overcome before it realizes its potential.
Alexandra Abate, Data Scientist at Dia&Co
Recruiting is probably the main one. Machine learning is still a fairly new area and as a result, there is a very small pool of candidates who have extensive real-world machine learning application experience. The harder problem, though, is that often companies don’t know exactly what they are looking for in a machine learning hire because they too are unfamiliar with their own machine learning application, and are unsure which technical and professional skills are most important to prioritise when screening candidates. This causes recruitment to be a slow and painful process on both sides of the equation.
Nikhil Aggarwal, FinTech Entrepreneur in Residence at iValley Innovation Center
Let us take the example of an ML Team which claims that they are using deep learning and artificial intelligence to reduce false positives by over 60%. Their pitch while brilliant is overly technical and complex. The core challenge here is that the ML Team has not bridged the 'gap' between risk management/business leadership and technologists. Often, the value proposition focuses on a predefined solution utilizing a particular machine learning technique which may not capture a broader set of evolving risk nuances. As a result, a proposed solution may check the technology boxes, but will not address all the underlying regulatory, compliance and operational risks.
To increase the adoption of machine learning, practitioners must explain their solutions and answer the fundamental questions on how core issues are being addressed. The solution must be implementable.
Haftan Eckholdt, Chief Data Science Officer at Plated
Operationalization. Most business practitioners have never met a modeler, much less used a model. No models can thrive until people know what modelers can do, and how to use models.
Michael Kubiske, Director, Center for Machine Learning NYC at Capital One
This problem is very different for different industries. Broadly, the labor market is one of the biggest impediments across the board. Companies that are less sophisticated in machine learning may not have the insight to hire the right person for the job, or they may be unwilling to spend the money to hire that talent. Apart from that, explainability in AI (essentially, unpacking the black box that decides how decisions are made—we have a whole work stream dedicated to this), regulatory concerns, and the ability to fully understand networks can all be impediments to further adoption of AI applications.
Saket Kumar, Chief Data Scientist at Google
There are a lot of positive trends (computing and storage costs going down). There are still data silos with and across organizations. One obvious one is the lack of qualified data scientists. Most companies - with the exception of large silicon valley companies - struggle to get right talent as the pool to draw from is not large.
Jérôme Selles, Director of Data Science at Turo
When it comes to applications of machine learning, the expectations are usually very high. The full lifecycle of a machine learning project is not necessarily well understood and that can drive disillusion within an organization and for the users. Depending on the quality of the data that is being used, automating the learning loop can be a challenge and, today, requires manual supervision. A good illustration of that is what happened with the Microsoft chatbot Tay that became racist within 24 hours. For machine learning to achieve its own potential, the learning process needs to be kept under control and values need to be respected. Data quality for the models is as important as education values in our society and we need more automated and systematic ways to make this happen.
Nikhil Garg, Software Engineering Manager at Quora
I think most would agree that the single biggest bottleneck for all machine learning is software engineering. We all collectively in the tech industry are still figuring out the best practices, tools, abstractions, and systems that can enable large organizations to innovate in machine learning at a huge data scale.