1 The statistical problem is only a proxy
2 Nature does not shuffle the examples. We shouldn't
3 Invariance across environments buys extrapolation powers
4 Invariance across environments is related to causation
5 Invariant representations enable invariance
Tag yourself - I'm #2
Great insights on learning from Leon Bottou
Some people know that there are many possible futures. And once you accept this possibility - well, then you want to play your part getting us onto the best timeline possible. If you're one of these people - wherever you are, whatever you do, however you view your talents - this is the place for you.