There’s a multitude of ways of doing AI. We can then obviously dive into technology and discuss the differences between machine learning, deep learning, artificial intelligence and big data – but wait. Despite my technology background I also must take a different approach sometimes. That’s why today we will cover the differences between consumer-driven AI (consumer AI) and industrial AI, taking a business point of view.
That is: AI which is being applied to the interaction patterns shown by people, usually in online transactions or based on online behaviour, versus AI used in industrial business processes like chemistry, underground research and car assembly. To name just a few.
Because when you use AI in order to optimize your business processes or to stir the market with the data currently on your shelf, it is of the utmost importance to realize which kind of AI you’re dealing with. The differences between both are quite large and can be essential to the success of your AI project.
All right, talked enough, time to address what I mean. It’s pretty much those aspects which can explain the differences:
- How well can you understand your data?
- How expensive is making a prediction?
- Where do you run your AI models?
- Can your AI model be a black box?
How well can you understand your data?
A website monitoring the behavior of its visitors yields very understandable data. After all, moving your mouse from the left to the right means moving your mouse, and a click is a click. Most variants of consumer AI, which is stuff like this, result in pretty understandable data. And with understandable data it is relatively straight-forward to train machine learning models. It’s a matter of finding an adequate architecture, training your model, and you’re done. All right, that’s a little bit too optimistic, perhaps – the model needs to be tested too (for discrimination, to give just one important example!) and needs to be made ready to be deployed for utilization at scale. But at its basics, it’s not really difficult. The data you’re trying to use is relatively accessible.
For industrual AI, the story is different. The image above shows a visualized dataset which is the output of a Ground Penetrating Radar. Such a radar emits radio waves into the ground and measures its echoes. Objects in the underground, such as gas pipelines or cables, cause disturbances which can be visualized into the radargram you see in the image.
Us non-experts understand that a mouse click is a mouse click, but it’s getting more and more difficult to determine whether the shape near the yellow arrows is a pipeline or a can with soda that happened to end up in the underground. Often, being able to discriminate between those requires years of training and even then, additional probing is often required.
Ground Penetrating Radar is the perfect example of applying industrial AI, for example by attempting to reduce the number of excavation damages. But at the same time, it becomes painstakingly evident that as a result of lower understandability of your data, and increased noise, training AI models becomes more difficult. The first large difference between pure consumer AI and pure industrial AI is thus the understandibility of your data.
How expensive is making a prediction?
A perfect example of consumer AI is a webshop which sells electronics. It contains a lot of smart functionalities. With machine learning, for example, the webshop can determine that people who are interested in a certain product will likely also buy another one. Although this is no longer the most spectacular occurrence of AI, it does however stipulate our second point: how expensive is making a prediction for your use case?
And with expensive, I do actually mean the financial consequences of a prediction made by an AI model. In the webshop scenario, the cost will be relatively low: if I’m not interested in the recommended product, I navigate to another website. That’s some bad news for the webshop as they lost a prospective customer, but it’s not a disaster, since the odds are that I’ll return to one of their web pages soon. This is especially the case when they use a large marketing effort, which is pretty much a commodity today. Consequently, the financial consequences of a prediction in this scenario, especially when they are wrong, are limited.
With industrial AI, this changes. Because what if I am the maintainer of a solar farm? Solarpholtaic panels are extremely vulnerable and need to be protected against the elements. A heavy hail storm targeting your PV panels is not that great, especially when the hail is big – which is not uncommon even for the Netherlands.
Or what if I created AI models that are used for monitoring the production of oil on an oil rig? The quality of my predictions must be extremely high. In this scenario, errors can be disastrous. When consumer AI is about a missed prospect, errors in predictions at the oil scenario can be potentially life threatening. If there is a substantial increase in pressure, measured by the sensors but not detected by the model, followeed by an explosion? We cannot oversee the consequences. But if my model triggers too often, and often for nothing? Then again there is a problem, since the production process must be brought to a halt every time, incurring financial costs too.
There thus exists a significant difference in price between making a prediction for a consumer AI and industrial AI scenario. Low costs, often less than a cent, are related to consumer AI scenarios. Industrial applications of artificial intelligence can bring to light risks so large that the implicit costs of a prediction can be hundreds to even thousands of Euros!
Where will you run the AI models?
When you created an AI model, you must run it somewhere. We also call this running the AI model in production.
In most cases, this happens by installing a model in a data center or by deploying it in a cloud instance. That’s the case for many examples of AI you run into: Facebook’s AI models run in the cloud, where data is synthesized and new predictions are made many times a second. Also the webshop we discussed before likely runs their models in the cloud, since it’s highly unpractical to run a model on the customer’s PC…for reasons of installing software on their computers.
Bringing new data from a computer to the cloud takes time. Generally speaking, the latency is limited, often some hundreds of milliseconds. But when you are in the field, such as with the oil platform and solar farm scenarios, you cannot simply assume that there is an internet connection nearby. In fact, when you install sensors in the field you are often dependent on wireless networks which can be substantially slower than the internet as we know it. What to do, then, with your continuous stream of measurements, which are registered but barely can be brought to the cloud for processing?
Although in the case of consumer AI this often goes well – after all, the consumer interacts with the B2C organization over a fast internet connection – industrial AI applications often cannot benefit from such fast connections. Communications speed can even be the bottleneck for employing AI in your projects. Because when I’m monitoring my oil pipeline and a dangerous anomaly occurs, fully automated action must be taken… always and immediately. It cannot be the case that the model is triggered by the disruption but that slow internet speeds result in so much delay that disaster strikes. With industrial AI, intelligence must be done differently: it thus often takes place near the edge, as we call it. Thus: in the field. The full name is Edge AI.
What are the consequences of this for deploying AI?
To start with, devices that are equipped with AI models do often not have access to substantial computational power. Since they are running in the field, on PCBs or specific tailor-made solutions, they cannot harness the power of the cloud. Although developments within the field of Edge AI are moving forward rapidly, it often means that many standard machine learning architectures (e.g. ConvNets) cannot be run properly. Especially with neural networks, interpreting webcam imagery can become a challenge in the field.
Luckily, those developments are indeed moving fast, because various private and public research institutes are creating embedded machine learning architectures which can run in the field more smoothly. We can also identify this trend by taking a look at the Gartner Hype Cycle for Emerging Technologies which tracks the development of various technologies over the year. Although deep learning has shown much during the last few years and is currently at the peak of the hype, Edge AI is still in its infancy.
Okay, summarizing for the third difference between consumer and industrial AI – an important difference is where you run your model, and the consequences of this choice.
Can my model be a black box?
The fourth difference between consumer AI and industrial AI is the extent to which one’s model can be a black box. Suppose that we’re looking at the webshop again. It does not matter much why the AI model reasons that people buying product X also find product Y interesting… given that the prediction is accurate. Although through various privacy related scandals the tendency is moving towards aiming to understand AI behavior, in most cases it is not strictly necessary to fully understand what’s happening inside an AI model.
Industrial AI once again changes the story. When we remind ourselves of the oil platform and the solar farm, it is indeed important to know why models produce certain predictions. A machine which indicates that production processes must be stopped? Nice, but no-one who listens to the machine unless it explains why it must happen. Until it’s too late, obviously, with all the unwanted consequences.
In short: where in the case of B2C AI a.k.a. consumer AI models can get away with predicting without reasoning about why, industrial applications of similar techniques come with an increased demand for the why. Much less black box! That’s a real challenge we will have to face over the upcoming years, because a large amount of AI models cannot even explain why they produced certain results. But luckily, also in this branch of research effort is translated in fresh ideas and small breakthroughs.
Not black and white, but a spectrum instead
In this blog, we saw four major differences between customer oriented and industrial oriented AI. We saw that with industrial AI, data is often less understandable to humans when comparing it to consumer AI. Making a prediction often happens in the field and comes with considerable risks tied to making wrong predictions. By consequence, industrial AI projects are often more complex and more cost intensive. Finally, industrial AI models often cannot be black boxes, because no human being will stop expensive production processes unless he or she understands what is wrong… because it’s the human being which will face the consequences when acting on wrong predictions.
Yet, in practice, the contrast between consumer and industrial AI is somewhat less black and white than we present it here. Every application and by consequence every AI driven project resides somewhere on the spectrum between industrial AI and consumer AI. Whether employing AI is worth it, how data must be prepared before models can be even trained, where the models must be installed and how colleagues will work with the models? Those are all questions that are dependent on where your AI project is located on this spectrum.
Through an innovation workshop and a feasibility study that is a product of this workshop, Aime can give common sense based advice about whether employing AI in your organization is (1) necessary and (2) worthwhile. This obviously includes a roadmap to achieve success quickly, if the answers to the two elements is yes. With our expertise in the area of business, information technology and artificial intelligence, we can augment the domain expertise present within your organization and really start using the data on your shelf. Please feel invited to get in touch if you wish to discuss your ideas. We service delicious ☕! Get in touch.
- Date - 11 February 2019