Thanks for this, this and the mechanistic article are concise and clear write-ups that I wish I had available when I got into neuroscience.
I was familiar with some of your papers but didn't realize how influenced you were by David Knill. I was a grad student in the department when he died, it was a big shock (and obviously a big loss to the department and the field).
As a student working on normative models, I indeed get criticism from neuroscientists that what I work on does not propose mechanisms. I think most understand what normative models do and do not offer, but they still seem to be viewed by some as insufficient in rigor or punting on the "hard" part. Wondering if you have thoughts/advice on this? Maybe we just have to be very clear on their contributions to neuroscience (e.g. specific novel predictions and proposed experiments).
Normative models provide great predictions. Mechanistic models are hard, and, quite honestly, usually impossible. Experiments are given carte blanche even if they get at neither. So I think we simply need to declare that normative models are important in itself. They provide the best known predictions for human behavior.
Also, to produce mechanistic you generally need to perturb a system. Which most people who pretend to do mechanistic neuroscience ignore.
So my suggestion is to just stick to our normative guns. Normative questions are some of the most exciting questions. And the idea that one of Tinbergen's questions is much more important than the other three makes little sense.
IMHO we should try to understand which neuroanatomical components participate in subserving the behavior well-explained by a normative model... ideally lots of behaviors. The brain reuses its overall structure across behaviors, with even single neurons generally participating in a variety of behaviors. Only by looking for what's reused (conserved) across different behaviors, and then understanding those behaviors in terms of a unified normative principle, can we then get down and ask how neuronal circuits implement an algorithm approximating that normative principle.
In the way we typically talk about "mechanistic" investigation in systems neuroscience, I don't think mechanistic investigation gets at algorithms or tests normative models.
Thanks for this, this and the mechanistic article are concise and clear write-ups that I wish I had available when I got into neuroscience.
I was familiar with some of your papers but didn't realize how influenced you were by David Knill. I was a grad student in the department when he died, it was a big shock (and obviously a big loss to the department and the field).
One of the cleanest thinkers I ever met. I hope to carry some of his ideas into the future.
As a student working on normative models, I indeed get criticism from neuroscientists that what I work on does not propose mechanisms. I think most understand what normative models do and do not offer, but they still seem to be viewed by some as insufficient in rigor or punting on the "hard" part. Wondering if you have thoughts/advice on this? Maybe we just have to be very clear on their contributions to neuroscience (e.g. specific novel predictions and proposed experiments).
Normative models provide great predictions. Mechanistic models are hard, and, quite honestly, usually impossible. Experiments are given carte blanche even if they get at neither. So I think we simply need to declare that normative models are important in itself. They provide the best known predictions for human behavior.
Also, to produce mechanistic you generally need to perturb a system. Which most people who pretend to do mechanistic neuroscience ignore.
So my suggestion is to just stick to our normative guns. Normative questions are some of the most exciting questions. And the idea that one of Tinbergen's questions is much more important than the other three makes little sense.
IMHO we should try to understand which neuroanatomical components participate in subserving the behavior well-explained by a normative model... ideally lots of behaviors. The brain reuses its overall structure across behaviors, with even single neurons generally participating in a variety of behaviors. Only by looking for what's reused (conserved) across different behaviors, and then understanding those behaviors in terms of a unified normative principle, can we then get down and ask how neuronal circuits implement an algorithm approximating that normative principle.
But would those not be mechanistic questions?
In the way we typically talk about "mechanistic" investigation in systems neuroscience, I don't think mechanistic investigation gets at algorithms or tests normative models.