It has been difficult to avoid hearing stories how artificial intelligence will change many industries in the next coming 5-10 years. There has been a lot of buzz around artificial intelligence, but when will we get something more than just chat bots and computers playing games? Could I do something with it, even if my company is not one of the five biggest tech companies in the software business?
I would like to tell you that there are real-world use cases that can be realized without super computers or an R&D budget big enough to run a small nation. We, at Spectral Engines, have built an industrially relevant application, and it’s actually available already today. In this blog I’ll give you my view on how to build viable AI applications with relatively small resources.
Artificial intelligence makes the difference
We use artificial intelligence to analyze material composition in our new NIRONE Scanner solution. The principle is surprisingly simple. We have built a sensor that measures the near infra-red spectrum from a material and then sends it over to a cloud service that runs a deep learning model which can deduce the composition of the material the scanner is measuring. Our customers can then access the data through a mobile app or their own data acquisition system. The result gets back to the customer in a split second. This is a huge contrast to traditional laboratory measurements. Our Scanner products are faster, cheaper, can operate continuously and automatically, and they can be mass-produced without individual model calibration. All of this is achieved with the help of AI.
The most valuable advantage of deep learning is that it enables the mass production of sensors. Traditionally the spectrometers are individually calibrated for the measurements. Deep learning models can learn the difference between variance caused by the devices, and variance due to changes in the measured material. Given enough data, it learns to see through the noise in measurements caused by these small variations in manufacturing and environmental conditions.
Another big advantage of AI is that it can learn anything a human can learn, or even more. Additionally, when it has achieved this, it will never get tired, it can be replicated, and it will not go into a strike nor require salary hikes. Though we are still very far from artificial generic intelligence, machine learning is already bringing additional value in many highly specialized applications. At our company, we have already built a deep learning model that has super human capabilities in material recognition from the infra-red spectrum. In this way, we have already reached the next level of material sensing in many industrial applications.
Reliable data is the king in machine learning
As the astute reader might already know, deep learning is fundamentally a statistical method that learns to infer correlations within data set. This sets some limitations to its applicability. First you need to have data of your problem, and the harder it is, the more data you need. In our case this means that we need to collect data from the problems we want to solve. Sometimes it is very expensive, and the opportunities to do so are rare. However, this is hardly a unique problem, and therefore others have already solved it. There are ways to generate new artificial data from the existing data, and there are ways to reduce the dimensionality of the problem. Those means reduce the complexity and the amount of data that is required to train an accurate model.
The point I want to stress the most is that your application will never be better than the data you have collected. Therefore, this is the part that you should spend most of your budget on. Once your service is running, make sure you also save the data your customers send to you, as it will most certainly become useful to you as you learn to leverage AI better. Reliable data is the king of the machine learning business.
Smart choice mobile app and web user interface technologies save a lot of development time and money
Model building is only one part of the product, as you still need to make the model available for the customer – however, there are relatively inexpensive ways to do just that.
Our choice was to use vendor independent open source tools to build an application programming interface (API) for the analysis. There are tools that are easier to deploy, e.g. Google’s CloudML, with which you need very little effort to establish an API end point that runs your machine learning code. The trade-off with these proprietary tools is that you are then stuck with your service provider. I honestly believe that if you are building something that will only get a couple of requests per minute, you are better off paying a service provider to run all of your infra. Dev-ops engineers, who you need for running your own infra, are also a significant cost, so I doubt that you would actually save money by rolling your own API. Plus, finding good software engineers is hard nowadays, even if you know what you are looking for.
On top of the API, customers are very likely to want an interface for browsing the results. There are many ways you can accomplish this with a relatively low budget. If your business is like most, you want a mobile app and a web user interface. Something I would highly recommend is a progressive web app, which serves your both needs. Especially if you don’t need peripherals like Bluetooth or to lesser extent USB in your mobile app. The progressive web apps are great for provisioning, as your updates come from a single source, your website. This means that you don’t rely on third party app stores and your users don’t need to do anything to update.
With these steps, I believe you can go off to build a simple and lovable product that will bring the benefits of AI to your customers with a relatively low development cost and a really short lead time. Be prepared to maintain it for a long time though, as sometimes the customers really fall in love with these AI applications.
Our company offers a ready-made platform for material sensing AI applications. So, if you are looking for a scalable, field tested, and cost-effective material sensing solution, please get in touch with our sales people. With Spectral Engines, you own your data and applications, we just host them and help you develop them.
By Pekka Röyttä, Principal Data Scientist, Spectral Engines