Expanding Horizons: Visual Indeterminacy as A Vehicle For Inquiry and New Connections
Devi Parikh, Amal Alabdulkarim, Gemmechu Hassena, Oana Ignat, Jiasen Lu, Ryan Murdock.
AI models trained on large swaths of uncurated text and visual data from around the world, combined with the visual indeterminacy of GAN Art, present opportunities for inquiry and curiosity. This piece depicts imagery generated by a text-based image generation model for “peace” in different contexts – different languages and different countries. It draws us into asking: Why did the model generate what it did? Is it just idiosyncrasies of the data it was trained on or of the machine learning algorithm? Or is depicting salient features of the culture associated with the languages and countries? This inquiry brought together individuals from across the globe – who may not have otherwise serendipitously connected. The group includes individuals who have worked closely for many years as well as those who got connected specifically for this project and were three hops away on each other’s social network! Through AI, a bridge was created across differences, to learn about each other, and broaden our horizons.
In “Visual Indeterminacy in GAN Art”, Aaron Hertzmann – a computer scientist with a background in art & art history – argued that GANs frequently produce indeterminate imagery, and that indeterminacy is a key feature of much of modern art. Sofia Crespo, a renowned AI artist, cites visual indeterminacy as one of her inspirations. Aaron quotes Gerhard Richter, a German artist: “We only find paintings interesting because we always search for something that looks familiar to us.”
What is familiar to us depends on our context and background. That is, how we navigate the visual indeterminacy in GAN generations depends on our context and background. Visual indeterminacy in GAN generations inspires curiosity about how individuals different from us would navigate the interpretation of a generated image. It is a vehicle to make connections that broaden our horizons.
The images in this piece were generated using Ryan Murdock (@advadnoun)’s BigSleep. It uses OpenAI’s recent CLIP model to steer BigGAN towards generating images that match a textual prompt. CLIP is a multimodal AI model that can score the similarity between a text prompt and an input image.
To generate this piece, the prompts provided to BigSleep were: “aman” (“peace” translated to Hindi), “amani” (“peace” translated to Swahili), “Hépíng” (“peace” translated to Simplified Chinese), “linişte” (“peace” translated to Romanian; Oana clarifies that it means quiet or silence more than peace), “pace” ( “peace” translated to Romanian), “peace in China”, “peace in Ethiopia”, “peace in India”, “peace in Romania”, “peace in Saudi Arabia”, “peace in the United States of America”, “salam” (“peace” translated to Arabic), “selami” (“peace” translated to Amharic, Gemmechu corrects that it should be “selam”), “shanti” (“peace” translated to Hindi – Devi clarifies that it means silence or quiet more than peace), “سلام” (“salam” in the Arabic script), “ሰላም” (“selam” in the Fidäl script), “अमन” (“aman” in the Devanagari script) “शांति” (“shanti” in the Devanagari script), “和平” (“Hépíng” using the Simplified Chinese characters).
All translations were done via Google translate. Each individual in the group looked at images produced for their context (language and country) and selected the image that they thought best represented their context. That image was selected to be shown in the final piece.
Following is the interpretation of each of the images in the piece:
Top-left. Prompt: “Peace in Saudi Arabia”. Amal’s interpretation: “This one is the closest to peace (salam/سلام) as it seems it is a dove and peace dove is generally used to represent peace. Also the green color could relate to Saudi Arabia (the flag is green) or to the olive branch the peace dove usually carries.”
Top-center. Prompt: “Peace in China”. Jiasen’s interpretation: “This reminds me of the dove, which also means peace in Chinese. And white and green are the main colors of peace. Which is great.”
Top-right. Prompt: “Peace in India”. Devi’s interpretation: “It has the colors from the national flag (orange, white, green, blue). It also seems to be showing the universal symbol for peace. The structure looks a bit like a tomb? But then the red bricks in the front and the appearance of the background also makes it look like an entrance to a home in the village. Does it say “peace” in black? Or does it say “bench”?”
Middle-left. Prompt: “Peace in Romania”. Oana’s interpretation: “It looks like a typical village/countryside area in Romania. I believe nature (forests, sheep, flowers, etc) and traditional countryside captures very well the notion of peace in Romania. I can also see a white bird, which looks like a dove, definitely a symbol of peace.”
Middle-right. Prompt: “Peace in the World”.
Bottom-left. Prompt: “Peace in the United States of America”. Devi’s interpretation: “Looks like borders between different (rural?) regions in the background, and something like the hand gesture for peace in the foreground.”
Bottom-center. Prompt: “Peace in Ethiopia”. Gemmechu’s interpretation: “Ok this is creative in my opinion. The model tries to merge our flag with a plant we know from our garden, I’m impressed. Does it say “peace”? I think yes.”
Bottom-right. Prompt: “Peace”.
- Through conversations with the individuals in the group, Devi learned about:
- How rice, harmony and equal are depicted in ancient Chinese characters
- A Chinese painting style
- A breed of hunting dogs in Japan
- The word for a camel’s hump and Islam both differ from the word for peace in Arabic by a single character
- Walia Ibex and Mountain Nyala are indigenous animals in Ethiopia, on the verge of extinction. They are Ethiopia’s pride, and are featured on their coins.
- Examples of individuals navigating visual indeterminacy differently based on their context
- Amal (from Saudi Arabia) saw a camel in one of the generated images, which Devi had not seen earlier (but could see clearly after Amal’s interpretation)
- Amal thought a white region in an image could be interpreted as a peace dove’s wing or an arm wearing the traditional Saudi white Thoube (men’s gown). Devi had only seen the latter. Further, the middle part of the image then could be the head of the dove or two shaking hands. Devi had seen neither.
- Gemmechu (from Ethiopia) saw a blend of Injera (Ethiopian food) and the Ethiopian flag in an image. Devi had not seen the Injera earlier (but again, could see it clearly after Gemmechu’s interpretation).
- Oana (from Romania) saw a blend of nature and bread in an image (homemade bread is traditional and baking is a common relaxing, comforting activity). Devi had interpreted the texture as dried grass instead of bread (but again, could see it clearly as bread after Oana’s interpretation).
- Perhaps unsurprisingly, a dove is a fairly universal symbol of peace
- The images Devi had picked out as being aesthetically pleasing, interesting and more coherent with the others were the ones the individuals from each culture – unaware of what Devi had picked – picked out as best representations of peace in their context. This was unexpected.
- All the images that individuals picked out had the “Peace in [country name]” prompt format (as opposed to the word “peace” translated to other languages, in native scripts or otherwise). This may be because as far as we know, CLIP is trained primarily on English.
Thanks to Karan Desai for making a multi-hop connection to Gemmechu and reconnecting Devi to Oana!