Machine learning in manufacturing – Hype vs Reality

All the buzz 

There is so much buzz about machine learning and AI today. With the estimated spend of 1.2 trillion to 2 trillion in supply chain and manufacturing, businesses are encouraged to use these new techniques to modernize their operations and remain competitive. We spent the last decade collecting data. Now is the time to use it right? However, there are few who can identify an area where they have used it to solve a quantifiable problem, successfully. 

We’ve learned that machine learning can help manufacturing with predictive analytics, preventative maintenance, demand forecasting, etc… How do we get down to doing it?

Hype vs Reality

Let’s briefly talk about Artificial Intelligence(AI) vs Machine Learning. AI is a more advanced form of machine learning where the machine uses large sets of data and changes the algorithm based on intelligence aka learning. That’s when we start training a machine to start thinking more like a human – This is no easy task though because human beings are said to be irrational in their decision making. Decades of psychological research has shown, most people (and probably most philosophers too) are pretty irrational in their decision-making. For example, in 2014 the Hong Kong based investment firm Deep Knowledge Ventures made headlines about bringing in an AI algorithm as a board member mostly as a veto mechanism. Today, the company no longer uses the algorithm because big strategy decisions are based on intuition says Brian Uzzi. Largely because we don’t have enough data to make the “more human” decision.

The majority of businesses today are using machine learning ( where the program does not morph an algorithm) to solve complex business problems. 

No magic bullets yet… 

The reality is that problem solving still matters. There are no magic bullets. We need to close the loop between business and technology by having technologists understand the business goals and outcomes so instead of picking the tools that are hyped, they will choose the right technology for the business.

Monica Rogati had an interesting article about the AI pyramid. She says, “Think of AI as the top of a pyramid of needs. Yes, self-actualization (AI) is great, but you first need food, water and shelter (data literacy, collection and infrastructure).” 

Step 1 . Defining the problem is 90% of the heavy lifting (Discovery)

As with any technology, this is the first step to a successful outcome. Define the problem we are trying to solve. A Discovery or Assessment is what most technology firms call the process of trying to isolate and validate the problem area. Make it measurable and try to add a price tag or cost to it. This helps validate the need for the tool. 

When Bob says this problem costs the company $100,000.00 and he can have it fixed for $80000.00, that’s an easier sell to Katie the CFO.

Step 2. Roadmap (Ideate) – Refer to AI pyramid

Once we have the problem defined, then comes the next step of where can we find the data? How do we collect it? What’s the hardware we use (if any)? How do we store it?

A large part of this portion is working with subject matter experts to validate the solutions against the defined problems. This stage involves low cost ideation where you use simple tools (whiteboard, sticky note, Miro) to communicate an idea across effectively.

Questions about where we store the data (on prem or in the cloud), security of that data, the cost related, can all be discussed at a high level.

Step 3. Software Implementation 

Most times this involves setting up applications that collect data. They can be connected devices, mobile devices or IoT devices. These devices then collect the data and push into a storage, either within premises or in the cloud. Data privacy and security come into play when implementing this layer.

The infrastructure for software development environments for developing, testing, staging and production are built for continuous integration and continuous deployment. (Full CI/CD pipelines). 

How do we transform the data? How do we manage the data integrity? How do we identify and remove the outliers? How do we start versioning the models (machine learning algorithms) used?

Step 4. Validate to drive trust and reliability 

In a world where everything is fast paced and we want to keep moving on to the next best thing, it’s easy to pay less attention to this step. Especially because it may seem mundane but it is crucial. 

How do we test and validate the models? It’s not an easy task to pay attention to detail. We need people who specialize in modeling and validating. I would say doing this right is very critical to building end user trust in the system.

The reliability of the data and the data model is central to the success of this project.

Wrap up

The pyramid is an important fixture in this article. Paying attention to the infrastructure and being able to collect the right data, matters. If we can shorten the time needed to make changes to the implementation of steps down the pyramid so we are gathering the most insightful data in the top, that means we are making real progress. 


I want to thank Delbert Cope, the CTO of Blue Newt Software for sharing some amazing stories, learnings and ideas for this article. This article features information from the September 2019 issue of Fortune’s article, Learning to Love the Bot – Managers Need to Understand A.I. Logic Before Using it as a Business Tool

The artwork used is Nemanja Rosic’s 2011 piece called Epilogue. Acrylic on alluminum.