Challenges with Implementing Machine Learning

With artificial intelligence and machine learning becoming more popular every year, it would seem clear that more and more companies would want to join the modernization of business and implement some kind of AI technology into their own practices. But despite its popularity, many businesses still either are not interested in or capable of building a solution that is both up to date and that accurately addresses the company’s needs.

In 2019, O’Reilly published an e-book detailing surveys they had conducted concerning AI adoption among businesses and the limiting factors that prevent those businesses from either building an AI platform or furthering their investment in it. Four main responses made up 80% of the reasoning for not expanding AI: a lack of data or issues with the quality of the data available, a lack of skilled people or inability to hire for the roles necessary to complete the plan, difficulty in identifying best or appropriate use cases, and a company culture that does not recognize the need for AI or believe it is a necessary tool. This section will detail how to go about solving these problems that arise in implementing AI, making your model both more feasible and more effective overall.

An expansive data set is the key to training effective machine learning programs. Like humans, machines need practice in order to be able to recognize patterns or make judgments. The benefit of training a machine to do these things is obviously that they can do them infinitely faster than any human could, but the trade-off is that they need to see many more examples of a basic task.

Take, for example, recognizing a motor vehicle. When you were a child, once you learned what a car was it probably didn’t take you too long to be able to say what was a car and what wasn’t. But a computer could be shown a picture of a car alongside a picture of the same car from a different angle and be unable to recognize them as the same. Or once you showed it enough angles of that car to be able to recognize it immediately, you could then show it a picture of a truck or a sedan, and it would be unable to recognize they still are both cars just like your first image. The amount of data needed then begins to increase exponentially, and this example is just for a simple image recognition program.
First, you should figure out what kind of data you already have and see if it fits the model you’re trying to build. Categorize the data in any way you can; structured vs unstructured is a good place to start, but there should be any number of ways you can make different classifications. Once you finish that, you should be able to see what’s missing.

Filling in the missing data is the trickier part of the process, but there are several different routes available. Purchasing data sets is always an option, and probably the best option disregarding cost, as any data set you buy will be complete, very rich, and you’ll know exactly what you’re going to get. There are also many public data sets out there that can be accessed by Google data set search or through data scraping techniques that you can set up using RPA. It will be easier to decide on how you should be filling in the cracks in your data set once you see what you need.

A second issue many companies run into is a lack of human resources or knowledge about data science. At this point, there is a lack of knowledge regarding AI and machine learning as a whole. Educating yourself to the possibilities of machine learning is a great first step (like reading this book!) but, unless you’re employed by a mega-corporation like Apple, Google, or Microsoft, odds are building a highly skilled data science team will be more trouble than it is worth.

The best way to start is by educating yourself and starting with the basics. Do research on some easy use cases, look at the big firms and see how they position themselves with artificial intelligence, and don’t be afraid to ask around.  In order to maximize any AI program, you need the right mix of technical knowledge and business understanding, combining the ability to develop the technology with a macro awareness of what it’s being designed to do. But finding people with those two skills is extremely difficult, as many machine learning programmers aren’t experienced with business solutions, and data scientists in general are still fairly rare. It’s difficult to know what kind of talent you’ll be bringing in, especially if you’re operating on a budget, but there are more options available every day as the market for artificial intelligence grows, and temporarily hiring an established data team also can give you the kick start you need.

The final two issues are related in that if you cannot find legitimate use cases for AI and machine learning in your business model, it will prove difficult securing organizational buy-in for a technology that there is no definite use for. Use cases will differ from industry to industry, but almost any business can benefit from a well-constructed machine learning initiative. Retailers can use AI to create more effective prices, streamline logistics, detect fraud, and predict demand for different products. Financial institutions can use it to improve security, creating more accurate and predictive credit scores, and underwriting and risk analysis. Healthcare providers can identify at-risk patients more quickly, more easily, and in real time, as well as improve diagnostics and optimize the cost of insurance products. Again, more in-depth research into common use cases and investigation into how the biggest companies use machine learning should lead to a better understanding of what specific use cases can apply to your business.

As well, many business leaders are hesitant to trust decisions made by artificial intelligence because there is no apparent process behind the machine’s decision and it can’t be explained in a way that a person could comprehend. Most humans have rules that they like to follow, in business or just in life, and the only rules a machine learning program abides by are the millions or billions of numbers that are fed into it that it uses to learn how to make decisions. It is much easier for some to feel good about an action if there is a logical path that leads to that action, rather than just believing it because a computer’s algorithm spits it out, no matter how well-trained or advanced the computer may be.