Machine Learning: From hype to real-world applications

Machine learning: real-world applications from hype

How to use emerging technology business value.

The hype is real. Artificial Intelligence (AI) and Machine Learning (ML) are all over the media, and everyone wants to get involved in the technology race. The progress that has been made over the last few years has been tremendous, and you've probably heard claims such as "AI is the new electricity" and "AI will revolutionize our society completely." I will not comment on those statements, but whether we can all safely agree, is it certainly very interesting about these technologies. However, all this attention raises the important question: can it really live up to the hype?

In some areas, technology caught on, and it even surpassed the hype. Within image recognition, for example, the task of identifying objects and extracting information from images, AI is now moving beyond human-level performance (for example, machines are becoming better than humans at recognizing objects and images).

When applied correctly in the right use-cases, AI solutions can provide great value to your business. As far as AI and ML are concerned, with all the attention and hype, it is very important to be clear about how and when these techniques should be used.

Referring to the “Gartner hype cycle for emerging technologies” in the figure above, we see that “AI / Deep Learning” exists everywhere above the “hype cycle”. The advantages are that there is a lot of focus on the possibilities that these technologies offer. But, at the same time, one needs to be aware that in some cases there is a mismatch between expectations and reality. This makes it even more important to go beyond the hype and how AI / ML and data analytics can be properly applied to solve business-relevant cases and provide real value.

AI Revolution: Now why?

Why do we hype around AI? AI and ML have been around for a long time, but some key factors describe how these technologies actually started to launch over a few years.
An important factor is, of course, the amount of data available. Access to huge amounts of data is a key element that makes ML so powerful. The amount of data available is increasing exponentially, both the number of recording devices and the connectivity between objects via the Internet (IoT).

Of course, accessing huge amounts of data is one thing. Being able to process these data to extract other useful data. Access to affordable and powerful computer resources is critical to progressing within AI and ML. As a visual example of the tremendous development in computing power, we look at the picture above. On the left, we have the image of the NEC Earth Simulator (parts), the world's fastest supercomputer about 1 year ago. While today, you can buy a normal gaming computer and essentially access the same amount of computer power energy.

Complementary elements of access to big data and huge computing power have made it possible to process data on a scale that was not possible a few years ago. With the added benefit of open access to much research by tech giants like Google and Facebook, we now have a much better starting point for solving new and interesting problems.

Cross-functional support to build better products

Go beyond the hype and really build a solution that provides real value to your business, there are many important ingredients to consider. The technical aspects of combining algorithms, computers, and data are one thing, but you need cross-functional support to build better solutions.

Domain Knowledge: Unilaterally, you need someone with domain knowledge to solve the problem. What are the possibilities and limitations within a system, and how can the solution be implemented in practice?

Data Science: You also need capabilities within data science, which includes everything related to analytics, facts, data processing, machine learning, artificial intelligence, and in-depth learning. Essentially, methods and techniques to extract and use formats and information in your data.

Software Engineering: Software engineering skills are key to building good data-driven solutions. This includes laying the groundwork for data cutting and processing through appropriate pipelines and managing access to data, and building functional and user-friendly software tools for end-users.

The essential thing is that you need to successfully combine all these ingredients in the order of good products.

What is machine learning?

Machine learning is one of the major sub-sectors of AI, and one where a lot of progress is currently being made to develop new and better solutions. Without adding to the technical details, it is necessary to extract valuable information from ML data. Data, in this context, can be anything from text, video, images, sound, sensor data, etc. Common ML solutions can range from models that analyze and classify images and videos, to monitor the “health status” of complex industrial equipment through sensors. Data or predict future sales for your business. (We will return to some concrete use case examples at the end of this article.)

In-depth study: applications from hype

It is important to keep in mind that deep learning is not magical, but based on the mathematical model that processes information. Images (after all, images are just a collection of numbers as seen by a computer).

The term "artificial neural network" comes from the relationship between how information is processed in the brain. ANN models do not attempt to mimic the human brain, but they are instinctively inspired by the mathematical model of "neurons", one of the building blocks of neuron information processing (hence the name neural network).

The stratified structure of these artificial neural networks (pictured above) is why we refer to this type of model as "deep learning". The deeper layers of the model (left part of the figure) usually yield information about sharp edges and basic shapes, while the upper layers (right side of the figure) find better and more detailed structures.

Learning these models to classify images correctly is a time consuming and calculating process. We present the model with examples of basically classified images correctly, and the model attempts to classify new images by itself. Initially, the predictions are only random guesses, but by showing the model a sufficient number of images, and correcting each time it makes a wrong guess, it will eventually learn to extract relevant information to classify them correctly.

Surprisingly, we don't have to tell the model what to look for - it learns these features from experience itself (which we call "machine learning").

Progress in this area has been astonishing in recent years, mainly due to access to huge data and cheap computing power. Performance for many tasks is now beyond the human level (for example, machines are becoming better than humans at identifying and classifying objects and images). It has, of course, a lot of interesting and useful apps that we'll get back to soon.

AI and creativity

We often think of "intelligence" as the ability to do things that we would normally associate with human intelligence. A similar action e.g. Be creative to produce art or music. Interestingly, even for these types of creative “human” tasks, we are now seeing AI performing amazingly well.

As an example of this, I show the example above where the AI ​​model was trained to produce images in the style of the famous Norwegian painter Edward Munch. Given a trained model, we can then use pictures of new scenes as “inspiration” for the “AI-artist” and mimic the human artist’s painting style. Works amazing, don't you think?

The example above is probably not very useful (albeit entertaining). However, it well illustrates the potential of the underlying technology to extract meaningful information from images, which in turn can be very useful for other applications.

Real-world applications of machine learning

To show the big span on the topics we work on, I’ve picked up a few examples of how ML can be used in real-world scenarios. These examples range from data analysis and ML use to head duty condition monitoring of industrial equipment to quality assurance and computer vision for various image recognition and object detection functions. You can read about our other fun projects here.

Sensor data use quality assurance

The above case is for a specific production company, like that. Involves the design of complex equipment or a process plant control. Generally, anything where the end product is a kind of production unit.

The simplified visual sensor data clearly shown in the figure above is to use all available information and highlight the relationship between measured variables and product quality. What's going on? Affects product quality? Given the current operating conditions, what is wrong with my product? Which part of the product is important to adapt to improve product quality? These are some of the questions that we can answer through the proper use of available data.

Condition monitoring

As concrete examples, we e.g. Consider using sensor data to monitor the condition of the compressor. In this case, we monitor the sensor data (temperature, pressure, vibration, ++) and the ML model knows what is "normal" based on the associated experience. The model constantly searches for discrepancies in recorded data and indicates impending failure/breakdown. If the behavior appears suspicious based on the measurement variable, the model may generate an alarm or initiate preventive measures.

Early warning of impending failure is of great value for condition-based maintenance and repair plans. The use of such solutions helps to avoid sudden breakdowns due to early warning which allows preventive measures and control of critical equipment. If you are interested in learning more about how ML can be used for condition monitoring, you can read the article below.

Image recognition for quality assurance

The picture above presents an example where this technique has been implemented within quality assurance. In this case, the model analyzes the images looking for certain defined features such as corrosion, damages, cracks, bad welds, etc. Probably not as entertaining as producing art, but definitely very useful for many industrial applications.

Image recognition in aquaculture

Another example is an object detection application for monitoring fish health on a fish farm. This is a task that is very challenging today, as fish are not primarily accessible for direct underwater observation. A new innovative solution is to use underwater cameras and AI solutions for automated image and video analysis. This allows you to extract important information about fish health and wellness in real-time. This is great news in both animal welfare and the fact that healthy fish are more beneficial for fish farm owners. Learn more about this project here

From technology to business value: the main takeaway

The main takeaway message is that one should always focus on what brings business value. Emerging technologies may certainly be able to achieve this, but focusing on the business side is critical to the success of new initiatives and projects.

Collaboration between different areas of expertise is key, to build better products and innovative solutions. The best solutions emerge when domain experts and software/analytics expertise help bring out the best in what emerging technologies can offer.

If you’re curious about how to get out of the hype in real-life applications, feel free to discuss how technology and software solutions can help address your business needs.