Journey through the layers of the mind (2015)
A dramatization of what’s happening inside the mind of an artificial neural network. Based on #deepdream #inceptionism by Google Researchers Alexander Mordvintsev, Christopher Olah, Mike Tyka.
I’m sure you’ve heard of or seen deepdream, and you all probably hate it, because everything mostly looks like puppy-slugs. I love it. But I should be clear it’s not the aesthetics that I love, its the algorithmically conceptual poetry. I wrote an article about this which you can find here, but very briefly:
This is a deep Convolutional (Artificial) Neural Network, trained to recognize images. It’s very loosely inspired by our own visual cortex, a network of specially arranged neurons, which ‘activate’ (or ‘fire’) when they recognizes specific image patterns. And these activations cascade through the network, across many ‘layers’, where neurons on each layer respond and activate based on activations happening on the layer below.
What deepdream does, is kind of visualize what’s happening inside the depths of the network. It’s amplifying the activities in those intermediate layers, which would normally stay hidden or ‘latent’, and deepdream makes them visible. It’s quite similar to us looking at a cloud or a Rorschach inkblot, and recognizing shapes, and then verbalising (or drawing) what we think we see. That in itself I find interesting.
But my favorite aspect of deepdream is, that when we look at these deepdream generated images, we might say “oh look a puppy-slug”, or “a bird-lizard”. But actually, there’s no such thing. There are no birds or lizards or puppies or slugs in these images. There are only bird-like, puppy-like, slug-like features. The deepdream network is recognizing bird-like, puppy-like features in the original image, so those associated neurons are firing, but weakly. But then deepdream amplifies those activations to produce these abstract images, which contain certain qualities and characteristics of what it thinks it can see. As a result we humans pickup on those same bird-like, puppy-like, slug-like features that the network found, and the relevant neurons in our brains fire. And we complete this recognition process by projecting those meanings back onto what is essentially structured noise, and calling them “puppy-slug” or “bird-lizard”. And that poetry is blowing my mind.
This is a summary, I’ve written more about these metaphorical and philosophical connections and biases at
And a bit more technical information (for the curious):