Artfinder CEO sees visual search as the next frontier

Artfinder works to lead the way in AI for non-photographic visual search

“A lot has gone into [visual search for] photography,” says Jonas Almgren, CEO of Artfinder over the phone from London, but he adds “art is different”.

When we caught up early last year, Almgren explained how his company aimed to be like a dating site to connect art buyers and artists, and in the process, to take the highbrow sting out of purchasing original art. The main thrust of this was achieved through improved personal recommendations via graph databases.

“Everyone’s taste is unique,” explains Almgren. “[So] we wrap the site around you.”

Since we last spoke Artfinder has continued to build on this technological base. This has included launching a Twitter chatbot called Emma – which uses facial recognition software most commonly used by investigators to catch criminals. In practice, it allows individuals to tweet any image – whether it’s a picture of a Van Gogh, a smartphone selfie, or a snap of the sunset on the way home – and Artfinder will suggest an affordable artwork in a similar style. It was a “way to test new technology” and get people on the platform in an “engaging and fun way,” says Almgren.

Most notably at the end of last year, the company raised of $2.2m in investor funding from venture capital firm Oxford Capital and William Tunstall-Pedoe  – the entrepreneur behind the Amazon acquired voice recognition app  which was used to create the Amazon Echo. The funding will be used to further improve the technology – with an emphasis on AI – and to build a base in the US.

“It is remarkable to see how AI has rolled out [recently],” says Almgren. “Yet the visual language is completely unexplored.” In future he hopes to see some really interesting results in this area just like Alexa has delivered for voice. Tunstall-Pedoe has been speaking to the tech team, he adds, but how he will advise and assist moving forward has not been formalised yet.

Chief Technology Officer, David Tilleyshort, who is also on the call, explains that tech developments will hinge on using machine learning to optimise the recommendation algorithm. Over the last year or so the team has been doing a lot of manual A/B testing, he says. Moving forward machine learning will allow Artfinder to streamline and improve that process.

Tilleyshort is keen to stress that Artfinder is not just interested in how well users relate to individual art works but also – perhaps more importantly – how users relate to artists.

“Half the excitement is falling in love with an artist,” says Almgren who believes this should be centre of the experience. To facilitate this he advises artists to put as much information on the platform as possible. This means not only adding multiple clear images of paintings – to aid image recognition – but also the “kinds of [personal] data artists don’t think are important”. The more data available the more ways people can “find affinity” with an artist, he explains.

While this sort of data may not feed into the recommender engine immediately the better and smarter the engine becomes the more likely this will become useful in future. So, if for an example, a potential buyer seems to favour brightly coloured abstract watercolours by artists from Kent, the engine will gradually show them more of these types of pictures. We’re “just scratching the surface” at present, says Almgren.

The use of AI will greatly speed up this process, and Tilleyshort adds that the team is also looking to understand the relationship between clusters of people who “in some way fit together but not in a binary way”. He describes this as attempting to catalogue the data in a similar way to how a social network, like Facebook, might.  

Technology developments aside, Artfinder is also looking to use funding to open its first US office in Miami and enlist new US-based artists. Almgren explains there are around 9,000 artists on the site in total – the majority are in the UK – and around 2,000 are in the US. The numbers are growing, he says, but artist numbers stateside are “a weakness” because it becomes expensive for users to buy paintings that need to be shipped internationally. 

Artfinder’s job is to make people realise that affordable art is out there, concludes Almgren. The facial recognition technology used was originally employed to find criminals “but it has been manually tuned”. Once you start automating that process you can begin to see the real power of AI to deliver very accurate recommendations.



Also read:
Artfinder: “We are in many ways OkCupid for art”
The next wave of disruption: Graph-based machine learning