AlphaGo has won its series in the game Go against grandmaster Lee Sedol 4-1. I wrote an initial post about AlphaGo after its first victory against a lesser ranked player. Humans have very big brains compared to the neural networks used by the program which shows that humans are unlikely to be able to use much of their brain for any one specific task. This, combined with the ability to run machine networks fast and against a lot of training data will make this technology formidable for many tasks.

Many people have been claiming that creativity will be one area in which machines will not be competitive with humans any time soon. But it is not clear that this is true. Creativity is related to the process of conjecture. Every new design, new text, new scientific theory, etc is a conjecture of a possible future state. The human brain is very good at coming up with such conjectures.

But here too we should notice something: if you want to come up with a new architectural design it helps to have learned a lot of existing designs. Einstein read a lot of the work of other physicists. Put differently the first step in creativity and conjecture seems to be observation and training of a network based on that.

And this is is the part where AlphaGo comes in again. Inside it are two neural networks: one that conjectures new moves and one that evaluates those moves. This split pattern could turn out to be applicable to many other domains and could give us “machine creation” (as opposed to “machine classification”) much faster than we are currently expecting.

For instance, you can imagine a network trained in a style of music that can be used to conjecture new notes and then another network evaluating whether the conjectured notes are a good direction. I would not be surprised to find such an approach producing high quality new music in the relatively near future. 

AlphaGo has won its series in the game Go against grandmaster Lee Sedol 4-1. I wrote an initial post about AlphaGo after its first victory against a lesser ranked player. Humans have very big brains compared to the neural networks used by the program which shows that humans are unlikely to be able to use much of their brain for any one specific task. This, combined with the ability to run machine networks fast and against a lot of training data will make this technology formidable for many tasks.

Many people have been claiming that creativity will be one area in which machines will not be competitive with humans any time soon. But it is not clear that this is true. Creativity is related to the process of conjecture. Every new design, new text, new scientific theory, etc is a conjecture of a possible future state. The human brain is very good at coming up with such conjectures.

But here too we should notice something: if you want to come up with a new architectural design it helps to have learned a lot of existing designs. Einstein read a lot of the work of other physicists. Put differently the first step in creativity and conjecture seems to be observation and training of a network based on that.

And this is is the part where AlphaGo comes in again. Inside it are two neural networks: one that conjectures new moves and one that evaluates those moves. This split pattern could turn out to be applicable to many other domains and could give us “machine creation” (as opposed to “machine classification”) much faster than we are currently expecting.

For instance, you can imagine a network trained in a style of music that can be used to conjecture new notes and then another network evaluating whether the conjectured notes are a good direction. I would not be surprised to find such an approach producing high quality new music in the relatively near future.