The deep learning techniques are here – and they are available to everyone
IT giants like Google are dominating the artificial intelligence market, but the good news is that they have also democratised the technologies so that they are now available to everyone. This creates completely new opportunities for Danish companies.
It is especially the development of deep learning in recent years that has started an industrial development. Deep learning is a machine learning technique that can be used for e.g. speech recognition and image analysis.
Democratisation of the techniques
The major breakthrough in the last 5-6 years is mainly due to the fact that a lot of annotated data has been generated by our smartphones and sharing images on social media. When data is annotated, it is a way of telling the computer that an image contains e.g. a cat or dog – and often also where in the image. But a lot has also happened in terms of algorithms and hardware, so computers have become faster and better at training neural networks. This has made it possible to use these techniques that are otherwise old.
The point comes from Henrik Pedersen, Team Leader in Visual Computing Lab at the Alexandra Institute, who currently collaborates with NVIDIA Deep Learning Institute (DLI) to run a number of hands-on courses on how to start using deep learning in computer vision. NVIDIA is one of the world’s largest manufacturers of GPU-based graphics cards, but today, artificial intelligence hardware is their primary business. Henrik Pedersen explains:
“Now there are databases with images in all categories. This is e.g. due to a sharing culture that has emerged with annotated images, thus making databases and turning the tools into an off-the-shelf product. For example, there are databases with people’s movements in connection with sports exercise, and then it is relatively easy to train a neural network that can analyse golf swings or dance steps.”
Easy to get started
Another important point is that the techniques do not require domain knowledge. You do not have to know a lot about computer vision. If you have images of dogs and cats, then there is an off-the-shelf product that can be trained to distinguish between dogs and cats. The new thing about deep learning is that you can start working with complex programs because the data is available.
“It is obvious if you are a company in possession of unique data. It may be microscopic images of cancer cells. Annotating them allows you to train a neural network to distinguish between cancer and non-cancer cells. If you have annotated your data by e.g. dividing them into appropriate categories, you can get started quickly.”
You may not get 100 percent recognition in the first instance, but according to Henrik Pedersen there are, however, some simple tricks that they teach in the course. The idea of the course is to avoid having too much theory and instead to get you started by teaching you how to work with the tools, so that you get to know quickly how you can train a neural network.
An obvious example of a company that has integrated deep learning in their business is Tattoodo. They have a social network where people put images of their tattoos. It gives them a lot of unique data that makes it possible to guess the style of the tattoos. This corresponds to the example of the dogs and cats. Here it is just a neural network in tattoos.
“A part of the story is also that the person behind has used the tool that we teach. Tattoodo was fundamentally challenged by spending a lot of time on classifying the tattoos but because they have been given access to unique data via their members, they can train an AI network to classify their images. At the same time, it can be used to suggest hashtags during upload and editing. In addition, they can provide inspiration to their members based on the motives they like, or based on the artists they follow.”
Keeping an eye on fertilised eggs
The deep learning technique can also be used to find things in images. This applies to everything – from humpbacks to cars. Typically, it takes a long time if you are searching through large amounts of images manually, but the process can be automated by training a neural network. The tools are available as off-the-shelf products and are ready to download and start using, explains Henrik Pedersen.
Many companies use the technique. One of them is Danish Vitrolife that makes incubators for fertility treatment. They keep an eye on the development of the fertilised egg by taking microscopic images. The procedure is extremely time-consuming, but by using deep learning a computer can keep an eye on the development.
Another example is RetinaLyze that works with retinal images. By looking at the retina, an expert can see e.g. the effects of diabetes that may cause blindness. When having an eye scan at an optician’s, obviously you do not have access to medical experts, and therefore work is on improving software that can screen for diseases.
Inspects the Great Belt Bridge for cracks
Deep learning can also be used to look at paint or concrete to see if there are damages and cracks and if repair is needed.
For a company like Sund & Bælt that is responsible for the maintenance of the Great Belt Bridge, it makes sense. Instead of having four men climbing on the outside of the bridge at 250 metres altitude for three weeks, they have started testing drones that can automatically identify the places that need repair. Deep learning also means that instead of reviewing 3,000 images, they get 100 qualified images of places that need to be repaired. Simply because the software has been fed with knowledge about concrete structures.
Another example is EIVA that e.g. develops software for collecting, visualising and processing data from sensors mounted on underwater robots. It can be used e.g. in connection with placing pipes on the seabed or placing offshore wind turbines. However, with several kilometres of pipes on the seabed it is extremely time-consuming to inspect for damage, and therefore they have been able to automate the process by feeding deep learning features in the software with images of pipes and the constructions.
Getting to know the surroundings
One final example is MIR or Mobile Industrial Robots that is part of the Funen robot cluster, and which has recently been bought by American Teradyne. They produce robots for e.g. hospitals. The hospitals have robots from different manufacturers that cannot communicate with each other. It happens so often that the robots block e.g. the doors and other narrow passages for each other.
“Deep learning allows you to train a neural network that can localise the surroundings. When a nurse comes on a scooter or other things happen that are unique to the surroundings, the system can recognise it. And again, it does not require domain knowledge. All you need is the data. Quite simply, we go to the hospital and make a form of video surveillance that records movements. Once this has been done sufficiently many times, the system has learned to recognise the surroundings.”
Henrik Pedersen does not doubt that deep learning techniques have a great potential.
“It is said that in a country like Germany, everybody who works with deep learning are getting into the car industry. Virtually all the parameters associated with self-propelled cars involve deep learning. That is to be able to understand the surroundings and to see that here is a bike rider and here is a pedestrian. We do not have any car manufacturers, but there are plenty of other options and we have a lot of data. This is especially true in healthcare, where you can take a cardiac scan, and by using deep learning you can extract some data, thus automating a radiologist’s work. There are a lot of people who are looking that way.”