Machine learning is already a big part of modern life. Machine learning algorithms work in Google’s search results, Spotify’s music recommendations, and Uber’s ride-sharing matches. The use of the technology will accelerate as companies find new applications in finance, health, energy, manufacturing and other sectors. Machine learning brings great opportunities for new businesses, but poses significant disruption to established companies and traditional employment. There’s much uncertainty but it’s clear that machine learning will be a transformational force over the careers of current MBA students. HBS’s CODE Club hosted a workshop in March to help MBA students understand the technology and how to succeed in machine learning.
How does it work?
Machine learning is when algorithms improve their performance at a specific task with experience. A simple example is predicting home prices. If you have a list of houses with different features – say, number of bedrooms, location, floor area – and their market price, and you would like to know how much a different house should cost, you could feed in the new house’s features and use the algorithm to predict the price. The more example houses you have, the more accurate the prediction will be. This is a classic example of ‘supervised learning,’ where the algorithm learns by example. Other kinds of machine learning include ‘reinforcement learning,’ where the algorithm learns by experimenting, e.g. teaching itself how to maximize trading returns; and ‘unsupervised learning,’ where the algorithm learns by observing patterns, e.g. grouping transactions to identify fraud.
How are companies using machine learning?
There are many applications for machine learning in the real world. Applications fall into two brackets of value creation: increasing customer service levels or capability and increasing efficiency.
Examples of increasing customer service or capability:
– Improving service: Netflix enhancing its quality of recommendations; Amazon’s Alexa recognizing complex requests faster and more accurately; UberEats predicting food arrival times
– Data handling and processing: the start-ups Forge.AI aggregating financial data to support trading and Talla creating an internal search function for corporate information
Examples of increasing efficiency:
– Increasing efficiency: DeepMind cutting energy usage at Google’s data centers
– Automating tasks: Financial institutions automating trading decisions
– Using predictive analytics: Industrial companies anticipating machinery failure and conducting pre-emptive maintenance
– Optimizing physical processes: Logistics firms finding more efficient shipping routes; planners managing traffic flows in cities
What do these applications have in common? Machine learning works especially well for problems that have large datasets, high levels of complexity and high associated value. There are two important trends. First, applications are moving from digital to physical. Industries with high levels of digitization will be changed by machine learning first. For instance, Goldman Sachs recently estimated that one programmer can replace four traders. Trading is a good target because it is all digital, has large existing datasets, and the definition of success is clear. But technology applications are moving rapidly towards less digitized, more physical industries. Even large industrial companies are getting involved: GE is gathering much more data on their industrial equipment and has developed a software platform called Predix that seeks to improve efficiency of operations and maintenance and incorporates machine learning.
Second, applications are moving from data analysis to controls. Data handling and analytics lie at the heart of many of the problems that machine learning companies have worked on to date. In the future, machine learning will play a greater role in controls. Controlling physical systems, e.g. warehouse automation, industrial equipment or electricity grids, has a far higher level of complexity but brings much higher value.
What are the different business models for machine learning?
Three business models are emerging:
1. Building new technology
Much of the new technology development is done by large tech companies or universities. This includes developing new programming languages, tools or algorithms. This area is hard for new companies to compete in. Building new tools takes a lot of resources and many of the existing languages and tools are open access, e.g. Google created its own software library for machine learning, called TensorFlow, which is widely-used and free. Many of the open source languages are not monetized but can encourage developers to use computing power on Amazon Web Services, Google Cloud and others.
2. Creating platforms
The business model for large tech companies has evolved from purely developing new technology to creating the platform for other machine learning companies to use. Amazon and Google host machine learning on their clouds, providing access to large computational power. Google Cloud AutoML and IBM’s Watson Data Platform allow users to run machine learning algorithms with their own datasets without the user writing any code. This opens machine learning to a much greater user base and could grow rapidly in future. These platform offerings make money by charging to host data, use cloud computing power and run machine learning tools. There’s a hidden benefit for companies with machine learning platforms – the more data people input, the more experience their machine learning algorithms have and the better their performance.
3. Opening new applications
The widespread development of programming tools and computational power in the first two business models means that other companies do not have to start development from scratch and can focus on applying existing technology to new industries. Sonic Vision, a startup in Boston, is working to improve accuracy of spinal anaesthetic injections. It uses machine learning and augmented reality to represent spinal scans in 3D. This company has adapted existing technology to a new application, used its specific industry knowledge and created an improved outcome for doctors and patients. There are many more opportunities for machine learning to be applied in new industries.
How to make the machines work for you
There are two elements to succeed in machine learning: data and people.
Data is essential for running and training machine learning algorithms. If companies can gain access to large datasets within a given industry, especially exclusive access, it will give them an edge over their competitors. The more data, the better the algorithm performs. Some of the companies that have been more effective in the space have developed creative ways for data acquisition. E.g. companies that sell energy storage and also install multiple sensors at the client site, which gives them access to customer energy usage data that the companies can use to improve their analytics services.
People with technical and specific industry expertise are also essential. Any aspiring startups must have technical expertise on their founding team. This is a key things that investors look for in a startup. Finding technical talent is no easy task: the market for developers is very competitive and skilled individuals can command high salaries. Individuals with specific industry expertise (sometimes called ‘domain’) are also important. This means someone with a detailed knowledge of the industry, key problems in that industry, a network and an ability to secure data. This is where MBAs can add value.
Any students who are interested in working in machine learning can start with these three steps:
1. Build technical understanding – a basic understanding of how the technology works will help identify a real solution and give credibility in interacting with technical people. Good introductions are Andrew Ng’s course on Machine Learning on Coursera and the textbook ‘Introduction to Statistical Learning.’
2. Find a technical cofounder – this won’t be easy: there’s a lot of competition and skilled programmers can command high salaries. Compensation is important, but there other factors can make an opportunity compelling e.g. working on challenging problems, global issues or new technological developments.
Find a compelling problem and build a deep understanding – the MBA can add value through building a deep understanding of a problem, acquiring the data and creating a viable business model for the specific industry.
Kieron Stopforth is a first-year MBA and Fulbright Scholar. He previously worked for Bloomberg New Energy Finance and interned at Google DeepMind. CODE Club is a community at HBS thinking about the importance of code in business – to join the conversation contact [email protected] Thanks to Ubiquity Ventures for sponsoring CODE’s Machine Learning workshop.