Last Saturday we all got invited to talk from Queens’ Computer Science graduate Demis Hassabis who matriculated here in 1994. Under the title ‘Understanding Intelligence’ he talked about his latest startup DeepMind (recently bought by Google), as well as what else he has been up to since graduating from Queens’.
Demis said he had always been interested in Artificial Intelligence. Before he came to Queens’ he created many computer games, most famously Theme Park (1994), which all had a strong foundation of AI built into them. He continued this after graduating from Queens’ with game titles such as Black & White (2001) and Republic: The Revolution (2003) until he in 2005 went to UCL to do a PhD in Cognitive Neuroscience. He told us that doing a degree in Neuroscience was crucial to achieve his in-depth understanding of how the brain works and how he later could use this knowledge to found DeepMind. He referred to the quote:
“What I cannot create, I do not understand.”
– Richard Feynman (1988)
In 2011 Demis co-founded a company called DeepMind with the idea to create ‘an Apollo programme for artificial intelligence’. The company was successful in attracting the best AI researches in the world and very impressively, only 3 years later, Google bought the company.
In difference to many others, instead of focussing on a very narrow task, DeepMind aimed to develop ‘General-purpose learning algorithms’ which would have a very broad range of applications. He outlined a sketch of how these works by showing the agent model:
- The agent receives perceptions from environment through a set of sensors
- The agent interprets these and decides on an what action to take
- The agent performs the action and goes back to (1)
While step 1 and 3 are easy to achieve, step 2 is where advanced artificial intelligence is needed and this is what DeepMind has mastered. Demis showed some demos of how their software was able to, only by watching the screen and being able to control the joystick, win any of the old Atari games. Because they had made their learning algorithm as general as possible, it didn’t really matter what games the software was playing, it would quickly learn what optimal strategies to use and become better than any human player. All they had to tell the algorithm was that a high score was good.
Demis finished his talk by answering questions on anything from self-driving cars to how a future with advanced artificial intelligence would look like.
I found the talk very interesting as I’m currently writing my final dissertation (Part II of the Computer Science course) within the field of AI and machine learning. You can read more about that project here.