Curation for Machines

CURATIon FOR HUMANS

The job of a curator for Humans is to be a subject specialist - To gain an overview of a certain topic and connecting the dots and lines by interpreting heritage material, historic timelines, prior art as well as past and current trends. This knowledge is often shared in the form of publications or exhibitions. Curators are able to pick specific objects to put them in context or to compliment an idea or concept. For publications and exhibitions curation is often about finding the unique most relevant and interesting object out of a vast amount of candidates. It is about the few or for the context of this text - A best of.

CURATIon FOR MACHINES

Curation for machines is different. That is because the goal is a different one. It is very often not about the few but the many that are very much a like. Part of the research material a curator would gather to substantiate his selection on needs to be reapplied and supplied when curating for machines.Taking the example of generating new images by utilizing a generative adversarial network, in short GAN, the aim of the curation process would be to define the topic or category and finding relevant material.Now there are two routes to go here. One could circumvent the curation process and throw a lot of broadly fitting data at the machine. However thinking of working with machine learning in an artistic context one might want to create a very specific effect or content. One might want to work with a partner that is striving to achieve the same goals. A supplied set of data is a representation and a road map of that goal. Proper curation and preparation of data is therefore a crucial part of the creative process.Curation for machines can be build upon a classical process of curation. A reduction of source material to find the best of, that one very special set of data or object, here an image, and then subsequently augmenting that set of data again within a narrow perimeter. Find or create look a likes, siblings and twins to deliver the context - the surrounding data. The special and unique has to become a mass product and in that sense it seems counter intuitive to the traditional curation.

DELETING IS EASIER AND QUICKER THAN FINDING AND SAVING

01On the search for the surrounding data it can be quicker to follow a naive approach of obtaining a large amount of broadly fitting data and subsequently thinning it out rather than performing a very precise and narrow search. Selecting and deleting unsuitable images from a folder takes less time than finding the right images online and manually saving each one of them.In the best case the pursued topic is already covered by a publicly available data set. A thorough search online often brings up data sets that can be used for a variety of purposes. Some of these data sets are listed here →

02https://www.cs.toronto.edu/~kriz/cifar.htmlhttp://www.image-net.orghttp…

03An original concept often requires an original data set. One good approach can be to create a data set from scratch only using original content. This can be a very long and tedious process though.Alternatively one could find suitable images which are publicly available on platforms such as flickr. Unfortunately this brings us back to searching and manually saving. This process can be automated though. This process is called web scraping. Web scraping is a very powerful and fast way of obtaining data that is available online. Depending on how it is done and what the source of data is this can be a bit of grey area legally which is something to be aware of. Read this a starting point.

04Once an initial set of data is obtained it often makes sense to determine specific criteria to quickly sort out unsuitable images.E.g. to sort out dark or night images from a set of landscape or webcam photos one could start by sorting by the creation date and time of the images and pick out the ones which have been taken after a threshold time. However the creation date and time might not reflect the moment the image was taken. Therefore another approach could be to simply go by file size. Dark images often show large areas of the same dark colour. In compressed image formats like jpeg this usually results in having to store less information and hence smaller file sizes compared to a colourful daylight scenery.

Purely technically speaking this would be a way to mass download images from flickr utilising their own search API. Full source code available on github.

def get_urls(image_tag,max_amount, key, secret):
   # using the flickr api library to get the urls for returned images
   
   flickr = FlickrAPI(key, secret)
   images = flickr.walk(text=image_tag, tag_mode='all', tags=image_tag, extras='url_o', per_page=50, sort='relevance')
   count = 0
   urls = []
   for image in images:
       if count < max_amount:
           count = count + 1
           print("Fetching url for image number {}".format(count))
           try:
               url = image.get('url_o')
               if url is not None:
                   urls.append(url)
               else:
                   print("Url for image number {} returned None".format(count))
           except:
               print("Url for image number {} could not be fetched".format(count))
       else:
           print("Done fetching urls, fetched {} urls out of {}".format(len(urls), max_amount))
           break
   return urls [...]

Curating a selection of cloud formations to teach the machine about the abstract shape of clouds, light and shadow, blue and white. 

Curating a selection of sunsets and skies accompanying Eric Cahans Sky Series to teach the machine about the gradients of light, sky and ground.