Recently Google published a blog post which showed what happened when they turned their image recognition algorithm on itself. The algorithm is normally used to recognise something in an image like a dog or a banana. But Google’s experiment repurposed the algorithm to not only highlight what it recognised but to amplify that aspect of the image and then repeat the process. The results are one ½ acid trip and ½ Van Gogh painting.
The experiment relies on using a computer system known as a Deep Neural Network which processes information similarly to neurons in the human brain. By using DNN (Deep Neural Network) Google hopes to make better sense of the vast reams of images online. Being able to recognise images will help them to roll out better search services and filter content more effectively.
Create your own deep dream here!
But as the Inceptionism experiment shows there is a problem with DNN based computer vision: a DNN will often make mistakes, recognising specific objects within an image which are not there and what’s worse is that it can be 99.9% certain it’s correct.
A paper by the Evolving artificial intelligence laboratory explores that a DNN (Deep Neural Network) can be presented with images that are ‘imperceptible to humans’ and ‘label the image as something else entirely’.
This also works the other way round meaning that malicious parties could trick image recognition algorithms for profit, inserting their images where they shouldn’t be. It may also prove problematic for proposed DNN security systems which could be tricked into recognising an intruder as a harmless photocopier.
Regardless of the problems differentiating between a golden retriever and digital noise, the applications for deep neural networks are vast. Currently there are many interesting applications such as self healing robots which when damaged can learn to function again by testing new ways to move and learning from their mistakes. It is this ability to learn dynamically is what makes DNN’s stand out compared to other methods of computer learning.
What Inceptionism shows us is there is an intriguing logic to the mistakes of Deep Dream. It is a fascinating insight into how deep neural networks are being utilised for new applications and arguably most excitingly creative experimentation.
Check out Google Inceptionism Gallery here.