Hofstadter and Dennett’s “Einstein’s brain” thought experiment is almost eerily prescient about today’s connectome enthusiasm: a complete wiring diagram may contain something essential about a mind, just as a record contains Beethoven, but staring at either one does not make the music happen. https://themindi.blogspot.com/2007/02/chapter-26-conversation-with-einsteins.html
I think here you expose exactly this distinction: the connectome constrains the computation, sometimes spectacularly, but leaves out synaptic strengths, time constants, delays, cellular physiology, plasticity, neuromodulation, and the history of interaction with the world. Once we supply all those missing pieces, however, we are no longer asking whether the connectome alone explains behavior; we are effectively building the “turntable” that turns anatomical structure into unfolding dynamics.
The Beat Saber fly is therefore less “we resurrected a fly from its wiring diagram” than “we discovered that a very complicated record makes interesting noises when connected to an amplifier, some software, and quite a lot of engineering.” The deeper lesson is that the connectome may be the score (or perhaps even the instrument) but explaining the animal requires explaining how the score gets played.
Does "fly sim" and "back-propagation" in the same sentence strike anyone as odd? Does anybody really think back-propagation is a biological function? Flies live 40-50 days, best case. Any knowledge that is not innate must be learned. How can flies survive their short lifetime if they can only learn by back-propagation?
Good point. In the post, I mention the distinction between biologically plausible RL (using the hardware in the simulated brain) and biologically implausible RL (using PPO or gradient descent based off of an external signal). It’s hard enough to make these sims work, I think using a biologically implausible method can be a starting point to get something to work, but an ideal simulation would use just the mechanisms in the sim fly brain to do its learning.
Excellent work, I was wondering about these fly sims. Thanks for breaking it down clearly.
This got me thinking about consciousness. In this case there's a huge disconnect between the input, the fly brain reservoir, and the output. At best, we'd have three different "minds" at play that aren't really communicating directly. The reservoir outputs some signals which the decoder then interprets based on the training.
There is a path where the input encoder could learn to encode better so that the fly brain can use it better. Then the decoder can make more informed decisions. In that sense they might merge into a single coherent system. But without allowing the fly brain to adapt, it's really not optimized to understand what it's doing. Maybe if the encoder understands required forms of the signals it can translate well enough.
Hofstadter and Dennett’s “Einstein’s brain” thought experiment is almost eerily prescient about today’s connectome enthusiasm: a complete wiring diagram may contain something essential about a mind, just as a record contains Beethoven, but staring at either one does not make the music happen. https://themindi.blogspot.com/2007/02/chapter-26-conversation-with-einsteins.html
I think here you expose exactly this distinction: the connectome constrains the computation, sometimes spectacularly, but leaves out synaptic strengths, time constants, delays, cellular physiology, plasticity, neuromodulation, and the history of interaction with the world. Once we supply all those missing pieces, however, we are no longer asking whether the connectome alone explains behavior; we are effectively building the “turntable” that turns anatomical structure into unfolding dynamics.
The Beat Saber fly is therefore less “we resurrected a fly from its wiring diagram” than “we discovered that a very complicated record makes interesting noises when connected to an amplifier, some software, and quite a lot of engineering.” The deeper lesson is that the connectome may be the score (or perhaps even the instrument) but explaining the animal requires explaining how the score gets played.
A good post. I'd been looking for something like it.
Does "fly sim" and "back-propagation" in the same sentence strike anyone as odd? Does anybody really think back-propagation is a biological function? Flies live 40-50 days, best case. Any knowledge that is not innate must be learned. How can flies survive their short lifetime if they can only learn by back-propagation?
Good point. In the post, I mention the distinction between biologically plausible RL (using the hardware in the simulated brain) and biologically implausible RL (using PPO or gradient descent based off of an external signal). It’s hard enough to make these sims work, I think using a biologically implausible method can be a starting point to get something to work, but an ideal simulation would use just the mechanisms in the sim fly brain to do its learning.
Excellent work, I was wondering about these fly sims. Thanks for breaking it down clearly.
This got me thinking about consciousness. In this case there's a huge disconnect between the input, the fly brain reservoir, and the output. At best, we'd have three different "minds" at play that aren't really communicating directly. The reservoir outputs some signals which the decoder then interprets based on the training.
There is a path where the input encoder could learn to encode better so that the fly brain can use it better. Then the decoder can make more informed decisions. In that sense they might merge into a single coherent system. But without allowing the fly brain to adapt, it's really not optimized to understand what it's doing. Maybe if the encoder understands required forms of the signals it can translate well enough.
Anyway, that's enough rambling from me.