When asking around to find the coolest startups in Helsinki, Curious AI was put forward. The company does not disappoint even its approach to being a startup is unique.

Curious AI is a research-based company founded 2015 as a spin-off from Aalto's Deep Learning Research Group. While conventional startups are trying to find profitable applications for the current state of technology, Curious AI is interested in something very different.

Other companies are using machine learning to solve problems now. We are driving the scientific development towards the next level of AI, explains Antti Rasmus, CTO of Curious AI.

Curious AI does cutting edge research, pushing the boundaries of the machine learning commonly called artificial intelligence today towards the illusive limit of true artificial intelligence. The proprietary patents the research leads to are the future incomes of Curious AI a business model similar to pharmas, with long risky development, but a big payoff at the end.

And like for biotech companies, there are some very big competitors trying to get there first. Curious AIs main competitors all have big academic and/or financial muscles. They are the tech giants all working on AI independently: Facebook, Googles Deepmind, Microsoft, etc. along with academic institutions and universities. So far, however, Curious AI is very much in the race.

Curious AIs results are state of the art.

Curious AI recently had a breakthrough, using unsupervised learning for object recognition in Google Street View.

What research team has come the furthest is difficult to say, considering everyone is working on different aspects of the problem, but Antti Rasmus does say this,

Our results are on par with the best in the world. Considering were a small independent group, its a pretty big achievement that Curious AIs is among the top three in the field.

To make this possible, Curious AI has a secret weapon in CEO Harri Valpola. A Finnish math prodigy, he was fascinated with brains from an early age and went on to study machine learning, theoretical neuroscience and artificial intelligence. Leading a research team at Aalto University led to Valpola co-founding ZenRobotics, which uses machine learning for robotic waste separation. So when he founded Curious AI to commercialise on the results of another research group he led, you could say hed been preparing for it pretty much his whole life.

The rest of Curious AIs team primarily consists of PhDs. It does, of course, also require substantial funding from global future-oriented VCs. Curious AI's current backers are Lifeline, Balderton and Invus.

Unsupervised learning is a major milestone for achieving true artificial intelligence.

he term artificial intelligence is thrown around a lot these days, but usually, when a startup says theyre applying AI to some problem, it just means they are using machine learning in varying degrees of sophistication. And there is a big difference. Not to say machine learning doesnt have huge potential for automation and optimizing computer responses to various problems, but artificial intelligence is at a completely different level.

A major obstacle to reaching artificial intelligence is solving unsupervised learning and this is what Curious AIs primary focus.

AI researchers have gotten very good at applying machine learning to the essential task of object recognition in images accuracy now actually exceeds humans. The problem is that usually a machine has to be trained with a huge set of images tagged with different categories in order to learn to recognize those objects in other images.

A human brain doesnt work like that, categories (concepts and objects) are formed without anyone pointing and saying Thats a cat, This is a dog. This is the goal of unsupervised machine learning, letting a computer form the categories itself as it learns to recognize them. Thats an essential task if computers are ever to be able to solve problems independently.

In one example, Curious AI used unsupervised learning to reduce the number of manual tags (nominations of objects in pictures) from 70,000 to 500 which saves quite a bit of work. For smaller companies that haven't been gathering their own data and labelling it for years or decades, any reduction in the amount of tags needed means lowering the threshold for applying macine learning, so it is extremely valuable.

Another big success of Curious AI is to use perceptual grouping - a way of clustering information. Other image recognition approaches may well be able to recognize that theres a car in a picture, but not be able to infer that there are in fact two cars. Curious AI clusters related pixels together, so that one car can be recognized as a separate object from another.

This also addresses the difficult problem of understanding layered images. It can be quite confusing for a computer to recognize a car obscured by a fence or tree so that only parts are showing. Curious AIs approach lets the computer infer that there is a fence in front of a car or two cars instead of recognizing just stripes of car. For self-driving technology, being able to infer that one object is behind another is of course invaluable.

This video shows what Curious AI are up to against the background of AI's historical development, have a look:

See the rest here:

This cutting edge AI startup from Finland is challenging tech giants and universities alike - Business Insider Nordic