Shortest possible syllabus for managers.
The basic principles of Machine Learning (ML) have not changed since the end of the 1980s. The trouble back then was that they couldn't be applied in practice, because the required computing power just wasn't there yet. Now it is. And it's reasonably cheap, too.
Graphic 1: Google trends, Machine Learning 2004–2019
Machine Learning can do things which are not cost effective for humans to do or are even downright impossible for mere mortals. People tend to under or over value correlations and are simply not capable of processing large masses of multidimensional data.
Graphic 2: Principles of machine learning (e.g. predictive marketing)
The greatest change to human life over the last four centuries has been the advancement of technology.
Graphic 3: Combining machine learning with automation
Workplaces have been transformed by the use of automation technology and automation has skyrocketed productivity.
When you combine automation with machine learning, the results can be quite amazing. For example, if you replace a man putting an object in a box with a robot, you can put a hundred more objects in boxes per day. But everyday, that productivity gain will be the same hundred objects. Combine that automation with machine learning, on the other hand, and you enable continuous improvement of productivity. In this scenario, the robot can use data on its own successes and failures to learn how to put objects in boxes even more quickly, so the productivity gain can grow over time.
If you want your machine learning project to succeed, you also need to be willing and able to change your processes and systems along with roles and attitudes. People who have long relied on their experiences and gut feelings have to learn to trust the algorithms. It can, of course, take time to demonstrate the reliability of the algorithms and gain trust, so communication and patience are important.
Further, analytics must be integrated with processes, and data utilisation must be comprehensive and applied from top to bottom. Otherwise, the impact of individual machine learning algorithms will be marginal.
Your machine learning applications might need capabilities such as understanding what is in an image or video; detect topics and sentiment from text; turn voice into text; etc. It doesn’t make sense to develop these yourself because they already exist, are quite feasible, and can be utilized via APIs.
These applications have already been developed by large cloud service companies with a continuous and large data flow related to the theme in question, a large number of world leading experts, and the necessary computing capacity. In other words, players like Google, Microsoft, and Amazon. For example, a voice-to-text service is available from Google and Amazon Sagemaker has a comprehensive portfolio of machine learning services.
If your machine learning need is business process specific, off-the-shelf APIs might not be applicable (although they can still be used to act as components of overall application). In this case you have to develop or source the machine learning model yourself.
What humans do better than machines, they understand business problems. In narrow-scope ML models the business analyst and ETL expert needs to extract data that closely reflects the real-life situation. The result is also affected by the multidimensional nature of the data set, and the expertise of the people modifying the data.
Analysts must understand, not only the data and algorithms involved, but also the business problem. The best results often come from team with multiple skills: a business analyst, a data extraction and transformation expert, and a machine learning (algorithm and coding) expert.
What's more, the analyst and the whole team must be as impartial as possible. Data must be allowed to do its work. An analyst is only an enabler, not an interpreter. Otherwise, the end result is art, not science.
When you use a large mass of data to teach and test a neural network, it’s not possible to “look under the hood” and see what it’s doing.
Statistical models, on the other hand – such as those using regression analysis and decision trees – can be explained, their results can be validated, and you can look behind the answer to see what’s going on.
Because of this distinction, it’s important to understand what your end requirements are when choosing the type of algorithm for your machine learning model. For example, under GDPR you might be asked to explain the logic behind your decision making.
A machine learning model follows the logic it has been taught. If it reads a hundred thousand blog posts about vaccinations being dangerous, it will oppose them. It has no ethical or moral bias. Without fail, it follows the content and values of the material used to teach it.
If you are using artificial intelligence to solve a problem, make sure that you are on solid ground in legal terms. If you reject a job application because of a recommendation made by artificial intelligence and the applicant sues you, what or who will answer the charges? ML cannot serve as justification in a court of law. And the GDPR will probably hamper the use of machine learning in marketing.
But the benefits of machine learning will be so great that companies will find ways of circumventing such problems. In other words, your competitors will use machine learning in their business.
But it's important to be aware of the risks.
There is plenty of talk at the moment about the kind of universal artificial intelligence that could solve a number of problems. Unfortunately, we have quite a way to go: general purpose AI is still years if not decades in the future. And, if you believe in Singularity, you probably wish this date was even further in the future.
In order for machine learning applications to work well, they must be defined as precisely as possible. Otherwise, the outcome will be of little value. That means they can be used only to solve a very precise and narrow problem.
Machine Learning will understand the content of images, create paintings in the style it was taught, listen to and compose music, play (and bluff) games, understand text, and detect anomalies in x rays. Machine learning will also optimise industrial processes, reduce hospital turnaround times, predict net sales, make sales and marketing process more intelligent, and assess lending risks.
Machine Learning will make a large range of activities more efficient. Many companies are still waiting to start their first ML projects. But when you ask somebody who has done even one project, they wish they had started way earlier. So there's no point in burying your head in the sand and saying that it doesn't concern us. It does. Maybe not today, but it eventually will.
Matti Airas is an expert in customer feedback management, marketing automation, predictive marketing analytics, and how to use data and machine learning to automatically trigger customer interactions. Before joining Tieto, Matti worked for a customer feedback analysis company Etuma and before that Nokia in the U.S.