My brother posted this article on Facebook a few weeks ago that asked,"Why is this viral image of unrecognizable objects so creepy?" TOTALLY LEGIT QUESTION. This
photo makes me pretty uncomfortable. The image was apparently created by artificial intelligence - specifically, a type that scans actual photos and creates new ones based on the data it has accumulated. In other words, this picture is a composite of many different images. So, colors, shapes, pieces of objects are familiar, but the pieces don't coalesce into an actual object that our brains can recognize and label. One of the main reasons I think this image looks "real" is the combination of colors and textures. In sort of the vaguest way, it looks like an old photo. It reminds me of photos in my grandparents' albums. If I keep the image in my peripheral vision, I can almost start to identify things sitting on a table with a white tablecloth: a stuffed animal? a calendar? a jar of something? some decorative knick-knacks? As soon as I start to focus on one particular object, it ceases to be identifiable. This is what causes my discomfort: We have evolved to be careful when dealing with unidentifiable things, because they might be harmful to us. Dr. Steven Schlozman, an assistant professor of psychiatry at Harvard Medical School, was quoted in the article: "I think the creepiness comes from our brains' attempts at recognizing a pattern, zeroing in on that pattern, and then having the expected pattern continually disrupted by another recognizable pattern."
This kind of artificial intelligence is called a Generative Adversarial Network, or GAN, and they basically create mashups from user inputs. You can create your own mashed up images with this GAN, Ganbreeder, and I did just that. There are some sample images that other users created on the site, and some of them are easily identifiable. The picture on the left, for example, is clearly a dog and flower. The picture on the right, however, is more difficult to interpret.
You have the ability to customize how much from each source image is used. I made the two pictures below. The first one is a candle combined with a beer bottle, and the second one is a snow leopard combined with a baseball (which, yes, makes me want to puke).
This left me wondering whether or not this kind of thing was happening with audio mashups and artificial intelligence, as well. The answer is...sort of. Or, yes, but with similarly unsatisfying results. I found something called Rave DJ, which allows you to input any songs (from YouTube or Spotify) and the site will mash them together. (This is very different from the Magic iPod, which I wrote about several years ago.)
You can pick as many songs as you want, but I started with just two: "Caroline No" by the Beach Boys and "Ride of the Valkyries" by Richard Wagner. The problem is...it's not exactly a mashup. None of the elements from either track are isolated, the the complete tracks are simply laid on top of one another and faded in and out. Have a listen here. It did create a mashup title (rather poorly): "Beach of the Boys" by Caroline No & Richard Wagner. In other words, artificial intelligence can't tell which set of words is the title and which is the artist, and it can't separate vocal and instrumental tracks from YouTube audio. It's hard to tell if anything is being done to either of these tracks in order to get them to "fit" better with each other.
I tried two songs that I KNOW mash well together with very little adjustment next: "California Gurls" by Katy Perry and "Tik Tok" by Ke$ha (see my blog post about these two songs and a resulting mashup). They are already in the same key and the same tempo, and also have essentially the same chord progression. The Rave DJ result was actually pretty awesome! Because of the similar features in the source songs, this one doesn't sound random at all. Listen here.
Finally, I used two songs in different keys that work well when mashed together (see this mashup by DJ Mighty Mike). The results were not great. While not exactly random, John Lennon and David Lee Roth are both singing lead vocals at the same time, and it's pretty confusing. The beats are matched, though, and the keys are shifted! The tempo of "Jump" was sped up, and the pitch of "Imagine" was raised a half step. Listen here. I think that this proves the point that DJ Earworm makes in his instructional manual, Audio Mashup Construction Kit. He basically implies that mashups need a human touch. Earworm spends quite a bit of time discussing key relationships between songs in mashups, but at the end, acknowledges that "it's best to let your ear confirm the choice" (99). This Rave DJ version of "Imagine" + "Jump" is perhaps the audio equivalent of the flower-dog created by Ganbreeder - you can tell what it is and how it was made, but it still doesn't seem quite right.