As for whether decision-making is ultimately about pattern recognition, my current take, informed by imbibing heavy doses of evolutionary epistemology literature written by biologists and ethologists (Konrad Lorenz, Rupert Riedl), is that yes, at some sufficiently low level, that’s all there is, and more complex systems including living organisms just wrap layers of abstraction and feedback around simple pattern recognizers.
I've been trying to work through the naturalistic decision making literature (starting with Newell and Simon, working our way to Klein and Gigerenzer). Not a lot of clarity, unfortunately. But this literature pretty clearly shows that we don't do the same things we program computers to do.
The distinction between "pattern recognition" and "optimizing models" is not obvious to me. Of course, at some intuitive level, binary classification feels different from LQR. But I think this glosses over the modelling that goes into setting up a classification task. Isn't picking the features, labels, and loss function also modelling? Maybe that's what you're getting at in the last paragraph here.
A related paradox (to me) is that learning a fixed predictive model is somehow seen as more general than doing state estimation, because state estimation requires specifying a model of dynamics. But the fixed predictive model also has a dynamics model -- one that says "nothing changes"!
Exactly! I expand on this point again today. I think the main distinction is that "pattern recognition" doesn't constantly articulate that there are models everywhere (features, measurements, stationarity, etc.). Whereas in optimal control-esque problems, the model is always explicit.
Thanks for the shoutout, Ben!
As for whether decision-making is ultimately about pattern recognition, my current take, informed by imbibing heavy doses of evolutionary epistemology literature written by biologists and ethologists (Konrad Lorenz, Rupert Riedl), is that yes, at some sufficiently low level, that’s all there is, and more complex systems including living organisms just wrap layers of abstraction and feedback around simple pattern recognizers.
I've been trying to work through the naturalistic decision making literature (starting with Newell and Simon, working our way to Klein and Gigerenzer). Not a lot of clarity, unfortunately. But this literature pretty clearly shows that we don't do the same things we program computers to do.
The distinction between "pattern recognition" and "optimizing models" is not obvious to me. Of course, at some intuitive level, binary classification feels different from LQR. But I think this glosses over the modelling that goes into setting up a classification task. Isn't picking the features, labels, and loss function also modelling? Maybe that's what you're getting at in the last paragraph here.
A related paradox (to me) is that learning a fixed predictive model is somehow seen as more general than doing state estimation, because state estimation requires specifying a model of dynamics. But the fixed predictive model also has a dynamics model -- one that says "nothing changes"!
Exactly! I expand on this point again today. I think the main distinction is that "pattern recognition" doesn't constantly articulate that there are models everywhere (features, measurements, stationarity, etc.). Whereas in optimal control-esque problems, the model is always explicit.