AI on LSD
What would you call an AI that is high on LSD and is a Fan of Pink Floyd?
No marks for guessing it right.. Its the Google DeepDream.
One might argue ‘what’s so cool about that?’. Given one can easily achieve such effects using image processing algos. The point to note here is that these results were achieved during unassisted learning. In simple words you can compare these to a toddler scribbling on a note pad; who can only write but not be able to read whats she has written. But the process of scribbling has begun and that is a big achievement in itself.
So how does AI create such fascinating experience?
Generative AI
Generative models are gathering a powerful momentum in the contemporary AI research. Leveraging the power of unsupervised learning and DL (Deep Learning), generative models such as Variational Auto-encoders (VAEs) and Generative Adversarial networks(GANs) create pictures, audio and video samples that look and sound more realistic with each training epoch. Among recent breakthroughs in the generative models, one should mention Sketch RNN, a neural net for generating sketches, Google’s DeepDream that creates surrealistic images by exciting certain neurons, and MIT Nightmare Machine that transfers horror style to images of human faces and buildings.
Future Prediction
One of the most exciting innovations in the field of generative models is a future prediction technique developed by MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL). The technique allows transforming still images into brief animations (GIFs) that simulate the future of the scene. The MIT’s algorithm has been able to produce short videos that look if as the artificial brain of the neural net was hallucinating. The imagery created by CSAIL’s neural net reminds of the tripping pictures that come to people on LSD. Motion blurs, unstable backgrounds and foregrounds one can see in the CSAIL experiment are also very reminiscent of the famous Surrealistic paintings by Salvador Dali.


The question arises: is this effect intentional?
On the one hand, machine ‘hallucinations’ may be explained by the shortcomings of the CSAIL’s generative model. On the other hand, they can say much about a hidden potential of machine creativity that expands the boundaries of human imagination. Let’s delve deeper into the technical details of the network to understand how MIT CSAIL algorithms produce this interesting effect.
Today we can’t make a head or tail out of it, call it spooky, hallucinatory or whatever but this is the cornerstone of the tomorrow.
AI/ML/DL got its due attention in the last decade and the results have started to show. Since the 70’s a oneliner was very popular amongst AI researchers and other techies
“Artificial intelligence can beat real stupidity”
That was then!
But in today’s Context we can say that ‘AI’s LSD can beat Real LSD’
Read More about the Architecture of the CSAIL Model
References:
[1] Open AI Blog. (2016). Generative Models. Open AI. Retrieved from https://blog.openai.com/generative-models/
[2] Vondrick, Carl, Pirsiavash, Hamed, Torralba, Antonio. (2016). Generating Videos From Scene Dynamic. Retrieved from http://carlvondrick.com/tinyvideo/
[3] Goodfellow, I. et al. (2014). Generative Adversarial Networks. https://arxiv.org/abs/1406.2661
