Wednesday, January 3, 2018

Beyond Literate Machines

Good thing I ran across this article today -
When A Machine Learning Algorithm Studied Fine Art Paintings, It Saw Things Art Historians Had Never Noticed
The Physics arXiv Blog via Medium, 2014.

Here's an article talking about a robot that can see similarities between artworks, and it's praised as finding something that no art historian has yet to discover. I wrote something about this on Network Address, because I like to write about more art-based things there. But there was a quote from the article that I thought was perfect for this blog in particular. They describe the process of training this algorithm to do its art-historian job. They feed it countless images, and with each one they have tagged with descriptions of its style, design, content, and (perhaps?) historical context. That way the algorithm 'knows' what it's looking at. And they conclude:

Comparing images is then a process of comparing the words that describe them, for which there are a number of well-established techniques.
Source document: Toward Automated Discovery of Artistic Influence, via Cornell, 2014

One of the main points of Hidden Scents is that the internet is and must be (for now) machine-readable - it must be made of words. Even the pictures must be reduced to words in order for this algorithm to 'see' them. Consequently, smell is word-averse. There is no language of smell, meaning there is no universal language to label the things we smell. Therefore, there cannot (for now) be such a thing as an internet for smells.

Image source: Sunmin Choi

Post Script
Causal diagrams by Edward Tuft

No comments:

Post a Comment