Neural networks hiding in waterfalls
As discussed in my blog “Creating AI self-portraits”, a generative neural network can be usefully visualized as a flow of connections cascading through layers of neurons. Below are some such visualizations to give a general idea:
With this concept in mind I fed DALL-E this prompt, “High quality realistic image of a deep neural network hiding in a waterfall”, to see how it would be interpreted. Here are some intriguing generations:
I find this image to be profoundly interesting, as it shows the AI dealing in metaphor. It has used the geology of the rocks to represent the layers in the neural network, while the hydrology of the cascading water is stylized to represent the “flow” of connections between neural layers (the descent gradient). In my opinion, this is a deeply creative interpretation of the prompt that gives a glimpse into the potentials I am looking for. The network is indeed “hiding”, but at a level of abstraction beyond mere simile.
.
.
This image is a bit more straightforward, consisting of essentially two parts. On the left we can see the AI representing itself similarly to the examples in “Creating AI self-portraits”, with a small amount of water splashing over it. This minor occlusion by water does convey the concept of hiding, but the concept does not dominate the image like in the previous generation. The right side is devoted to a more or less realistic rendition of a waterfall, making the overall effect of the image more about a mashup than a hiding.
.
.
This image deals with hiding in a somewhat different way. Like the in the first image, the cascading water is stylized to impart the idea of a flow in the connections between network layers. The layers themselves show up on the far left as very different rock morphology, but of course in this image the layers and the cascading connections are not overlaid, reducing the effectiveness of the metaphor. Hopefully, these partial solutions can give us insight into the associative processes at work in generating the images.
.
.
In this final image, the notion of cascading connections though network layers is abstracted to an almost literal waterfall. If we squint our eyes, as it were, we can imagine the layers of rock as layers in a structured neural network, but the water wants to read mostly as just water—making the metaphor either deeper or weaker, depending on how one views it. The concept of hiding is conveyed by having the metaphorical AI datafall become part of the main waterfall, but only after issuing from a completely artificial opening in the rock face.
.
.
So we have four generations that present different solutions to the idea inherent in the prompt. I am especially pleased with the first and last images, as they show the AI engaging in an adaptive and constructive way with some fairly abstract concepts—a true measure of creativity for me. The other two demonstrate a more limited “understanding” of those concepts and a generally weaker interpretation of the prompt.
And as is usually the case in generative AI, there were other solutions that show less “understanding” of the situation. In these images the concepts of waterfall, neural network and hiding are conveyed variously, but in none do the elements come together to form a satisfying interpretation of the prompt.
