When Alpha Go beat Lee Sedol, the world Go champion, it came up with creative new moves never previously seen before and even invented a whole new style of play unknown to humans. IBM's Deep Blue, the champion chess algorithm, failed to do either of these. What was the difference?

In this podcast, we review Alpha Go the Movie. Warning: Spoilers abound! Please go watch the movie first! This is an excellent movie. 

Bruce (using his admittedly thin knowledge of reinforcement learning) explains how Alpha Go works (using the materials previously discussed in our Reinforcement Learning episode) and how Alpha Go came up with a creative new approach to Go that went beyond the knowledge of the programmers. 

While Alpha Go definitely does not have "creativity" in the universal explainer sense of the word (it has no explanatory knowledge nor understanding), it did come up with a creative new playstyle never before seen in the history of the world that changed how humans play Go. Even the programmers were caught off guard by what it came up with. We talk about how Alpha Go challenges the Pseudo-Deutsch Theory of Knowledge but meshes well with Campbell's evolutionary epistemology. 

We wrap up our discussion with Sadia Naeem covering possible solutions and criticisms of those solutions. 

Sadia Naeem continues the discussion about unsolved problems in physics. This time we talk about (among many other things) symmetry and novelty. 

Sadia Naeem joins us again, this time to explain clocks, block universes, and eternalism. 

Sadia Naeem joins us to discuss her own research and musings into the problems and mystery presented by time.

In this episode Cameo, Tracy, and Bruce geek out over how good the Marvel TV shows are and how much they really get right. Spoilers abound, so be warned. 

Reinforcement Learning is a machine learning algorithm that is a 'general purpose learner' (with certain important caveats). It generated a lot of excitement with its stunning victory of Alpha Go against Lee Sedol the world Go champion.   

In this podcast, we go over the theory of reinforcement learning and how it works to solve any Markov Decision Problem (MDP).   

This episode will be particularly useful for Georgia Tech OMSCS students taking classes that deal with Reinforcement Learning (ML4T, ML, RL) as we briefly explain the mathematics of how it works and show some simple examples.  

This episode is best when watched on the Youtube channel, though we'll release an audio version as well. But the visuals are helpful here. The audio version is abbreviated and removes the mathematical theory and proof.

In this episode, we interview Chiara Marletto about her recent book The Science of Can and Can't: A Physicist's Journey Through the Land of Counterfactuals as well as discussing Constructor Theory in general and how it might help us form a new mode of explanation in physics. We ask her some tough questions about constructor theory and she fields the questions very well. 

For those interested in q-numbers vs real numbers, see Sam Kupyer's lecture on our Youtube channel.

Follow us on Twitter.

Check out our blog.

Donald Campbell made the bold prediction that all expansions of knowledge will be found to require the Universal Darwinism algorithm of variation and selection. In this episode, we're going to test that prediction and see if it holds up against what we currently know about Artificial Intelligence and Machine Learning.   

For example, does (apparent) knowledge created by Gradient Descent require variation and selection? Or is it really and truly inductive? Or does it just fail to create knowledge at all despite clearly creating improvements?   

Ultimately, we'll find that Machine Learning creates an exciting set of epistemological problems that need to be solved!

Youtube version with optional visuals

In The Beginning of Infinity, David Deutsch claimed that no existing evolutionary algorithm has yet created knowledge. But Karl Popper and Donald Campbell beg to differ and have argued that knowledge-creating evolutionary algorithms are ubiquitous, common, and easy to implement. Who is right? 

In this episode, we look at both arguments and assess them using Popper's epistemology. And along the way, we'll define the minimum requirements for an evolutionary algorithm (aka Universal Darwinism), explore what knowledge-creation is and then, finally, we'll attempt to answer the question of whether or not existing Artificial Intelligence algorithms create knowledge.

Visuals are available on Youtube. This episode may benefit from seeing the visuals. 

For more information on Donald Campbell's theory, this blog post covers his arguments. 

A summary of the contradiction between David Deutsch's argument and Campbell/Popper's argument is in this blog post.