I was excited to finally be able to understand what is meant by the term DEEP LEARNING on Wednesday evening. Terrence Sejnowski gave the Merson lecture originally titled A DEEP DIVE INTO DEEP LEARNING which he cheekily changed to A DIP INTO DEEP DIVING while the slide was up on the screen. How merciful that decision to take a dip was - by the end of the lecture my head was swimming with information on the history of artificial intelligence, how we are currently using it and possibilities for the future.
Terry is a computational neuroscientist based in the US who is part of Barack Obama's BRAIN Initiative. He speaks quite quickly and you can't help but imagine it is an attempt to keep the large amount of knowledge he has from spilling out. To chart the progress of artificial intelligence Terry used the example of facial recognition - something that humans are usually very skilled at, due to the visual cortex taking up over half the brain processes. It has taken a combination of learnings from engineering and neuroscience to come up with the various algorithms for a computer to recognise faces effectively.
I would describe Deep Learning as the process of taking what we know about how the human brain works and applying it to computers. Some of the features that have progressed artificial intelligence from neuroscience have been:
- recurring learning
- forming networks
- classical conditioning
- temporal difference learning
- reward systems
You might notice these advances on Facebook where an algorithm can not only recognise a picture but also caption it with some degree of accuracy. The latest advance has computers detecting human microexpressions which potentially gives companies the power to know how we really feel about their offer. Rather than be creeped out by this Terry notes that every computers have seemingly outstripped us, we have the ability to learn from them and become more skillful in turn. He sites the democratisation of chess as a perfect example.
Emphasising that each improvement is due to our advancing knowledge of how the brain functions, Terry left us with a utopian vision of the future classroom. A class robot will record each individual students progress, then create a learning plan tailored to their own rate and style of learning. I hope to see this in action if grandchildren eventuate, what a great response to Eleanor Roosevelt's entreaty. "What we must learn to do is to create unbreakable bonds between the sciences and humanities."