Another reason why algorithmic similarity would be useful is related to a recent line of thinking that I’ve explored recently. Specifically, the question is how we could regularize neural networks in order to make their computations more interpretable. The reason why a theory of algorithmic similarity would help is because we could apply some penalty to a neural network whose internal operations are too dissimilar to some understandable algorithm. This would encourage the neural network to mirror an interpretable computation which makes it easier for us to look inside and see what it’s doing.
Ideally, this would provide us the performance gains of neural networks while keeping the interpretability of GOFAI algorithms, like tree search.
Another reason why algorithmic similarity would be useful is related to a recent line of thinking that I’ve explored recently. Specifically, the question is how we could regularize neural networks in order to make their computations more interpretable. The reason why a theory of algorithmic similarity would help is because we could apply some penalty to a neural network whose internal operations are too dissimilar to some understandable algorithm. This would encourage the neural network to mirror an interpretable computation which makes it easier for us to look inside and see what it’s doing.
Ideally, this would provide us the performance gains of neural networks while keeping the interpretability of GOFAI algorithms, like tree search.