Computers vs connoisseurship

Could a computer program tell the difference? Left to right:
‘Portrait of Rembrandt with Overshadowed Eyes’, studio
of Rembrandt (c1629); Self-portrait by Rembrandt (1628-29)

The Financial Times ran an interesting article on computer analysis and identification of art works and paintings. The article is not technical in nature, but indicates that with more technology and more works being cataloged digitally, there is more effort and interest in using computers to analyze and assist in the authentication process.

The key appears to be the amount of useful and correct information that is compiled and can be analyzed.  So the old computer saying, garbage in, garbage out certainly could apply.

The Financial Times reports

It was towards the end of a fruitless day looking at auction catalogues online that a portrait of a fat, uniformed man in a hat flashed up on my screen. “European School, Portrait of a Gentleman, estimate $100-$200,” said the lot description. The picture, I thought, was perhaps late 18th- or early 19th-century, and the subject probably British. Was he some famous admiral?

I’d looked at hundreds of photos that day in my hunt for mis-catalogued paintings – called “sleepers” in the art trade – and by then was feeling too lazy to flip through my mental Rolodex of potentially matching sitters. So I idly copied the photo’s URL and ran an image search on a website called TinEye.com.

A few seconds later, I knew that a portrait of King George III lay waiting to be bought for a song in a minor US auction house. TinEye had matched it with other copies of the painting already online (royal portraits were frequently copied). My then employer, the art dealer Philip Mould, was delighted. I didn’t tell him I cheated.

I was reminded of this shameful episode when I heard of some new research published by computer scientists at Rutgers, the State University of New Jersey. They had developed a program which, they said, revealed how artists were influenced by earlier paintings through recognising (as tineye did) certain aspects of composition and technique. “Will computers,” the news headlines asked, “put art historians out of work?”

The program’s results were patchy, but not unimpressive. Its conclusion that Bellini was influenced by Titian was a little off beam, since it was the other way round (Titian was Bellini’s pupil). But Rubens certainly was influenced by Titian, as the computer argued, making many copies of his works.

I would argue that the Rutgers team began by asking the wrong question. In an episode of the 1960s TV series The Prisoner, Patrick McGoohan is confronted with a “wonder machine” that knows the answer to every question in the world, and will render man redundant. When McGoohan simply asks it “why?”, it explodes. Only humans can really begin to tell us why Bellini painted as he did.

But how Bellini painted is a more straightforward question, and computer science can help us with that. Like any academic discipline, art history is essentially about the accumulation of data, and art historians are coming to realise that computers let us compile, access and analyse information about art in an unprecedented way.

For example, a BBC website called Your Paintings has placed every oil painting in UK public ownership (some 220,000) online. At first, the data set was quite limited – artist, subject, date – but new information is being added all the time. Through the process of “tagging”, where objects within paintings from dogs to jewellery are identified and recorded, volunteers working with Oxford university’s Visual Geometry Group have helped train a computer program to recognise these objects automatically.

The accumulation of such data means we will eventually be able to ask Oxford’s program questions as varied as: “when did earrings first appear in portraits?”, “what was the rate of increase in female sitters in portraits over time?” and, as we introduce more technical information on pigments and media, “over what period did painting in oils replace tempera?”. The answers will give us quantifiable facts previously unavailable to art historians.

But what about the “who” of art: can a computer tell us who painted a picture? The term for this ability is “connoisseurship”, derived from the Latin cognoscere: to get to know or to recognise. Spend enough time getting to know an artist’s style, the theory goes, and soon you will be able to tell whether they painted a picture simply by looking at it.

Some art historians refuse to believe in connoisseurship. Others practise it enthusiastically, but are hopeless. A few get it right most of the time, which is perhaps the best we can hope for. It’s a fairly haphazard cognitive process, relying mainly on a lightbulb moment of recognition, and is of course fallible to all the flaws of human thinking: overconfidence (that you just know it’s a Rembrandt), or jealousy (that someone else spotted it when you didn’t), to name just two.

Surely a computer, not subject to such whims, could do better? If it was fed enough information on every securely identified Rembrandt in the world, and was able to recognise Rembrandt’s trademark brushstrokes, even the way he mixed his paint and ground his pigments, might it then be able to recognise his technique in other, previously unknown works?

Actually, no, and here’s just one reason why not. Pupils in Rembrandt’s studio regularly made skilled copies of his paintings, using exactly the same techniques and the same paints as the master. By every measurable metric these would be identified by a computer as Rembrandts. But a good connoisseur would tell you they were copies.

Great paintings speak to our soul not because they were made with a certain mix of pigments, but because they have been lit by the spark of human genius. Happily for jobbing art historians like me, that individual genius is something no computer can recognise. If anyone says they can, they’re cheating.

Source: The Financial Times