Just to reply briefly, the post was not meant to suggest that the edges and nodes alone are useful. The numbers are just used as an easy way to track progress, since most connectomics papers that actually attempt connectomic mapping report them. To do any useful or interesting simulation, you also need some idea of the strength of the synaptic connections, which can be crudely estimated by the size of the synapse or by more sophisticated means. It would also be nice to know what type of neuron we're looking at (some neurons specialize in a particular neurotransmitter, such as dopaminergic neurons). People are Janelia working on AI systems that can classify neurons by what type they are. It's also possible the shape of dendric spines and dendites may be important for subthreshold dendritic computations, glial cells are important, etc etc.
Isn't it reasonable to assume that we would only need to scan parts of the brain and then extrapolate to model a full brain? Especially as AI improves this seems reasonable. Have there been any generative connectomics models that generate plausible brains?
Also, do we currently have "physically realistic models" of nematode worm behaviour? Seems like that would be a first step.
Hi, just to respond.. AI might be able to reconstruct, to some degree, synapses & connections that are missing due to low resolution or artifacts in the scans, but I don't think you'll be able to get away with just scanning part of the brain.
Openworm has a physically realistic model and they have simulated worm motion (somehow by wiggling their bodies they propel themselves foward!). The simulation uses fluid mechanics (some version of Navier-Stokes) which is quite an impressive feat. However, to my knowledge nobody has created a worm model that fully captures actual worm behavior fully.
I think beyond mouse we will have to explore ways to upscale lower resolution images to higher resolution images using deep learning something like this https://arxiv.org/abs/2103.04989 or this https://www.youtube.com/watch?v=WCAF3PNEc_c&t . But for that first we need sufficient high resolution data to train the algos.
Zeiss will make 20k beam SEM which will suffice to get whole mouse connectome done in under a year.
This might be a dumb question but what do you do once you’ve mapped out these edges and nodes?
Just to reply briefly, the post was not meant to suggest that the edges and nodes alone are useful. The numbers are just used as an easy way to track progress, since most connectomics papers that actually attempt connectomic mapping report them. To do any useful or interesting simulation, you also need some idea of the strength of the synaptic connections, which can be crudely estimated by the size of the synapse or by more sophisticated means. It would also be nice to know what type of neuron we're looking at (some neurons specialize in a particular neurotransmitter, such as dopaminergic neurons). People are Janelia working on AI systems that can classify neurons by what type they are. It's also possible the shape of dendric spines and dendites may be important for subthreshold dendritic computations, glial cells are important, etc etc.
Isn't it reasonable to assume that we would only need to scan parts of the brain and then extrapolate to model a full brain? Especially as AI improves this seems reasonable. Have there been any generative connectomics models that generate plausible brains?
Also, do we currently have "physically realistic models" of nematode worm behaviour? Seems like that would be a first step.
Hi, just to respond.. AI might be able to reconstruct, to some degree, synapses & connections that are missing due to low resolution or artifacts in the scans, but I don't think you'll be able to get away with just scanning part of the brain.
Openworm has a physically realistic model and they have simulated worm motion (somehow by wiggling their bodies they propel themselves foward!). The simulation uses fluid mechanics (some version of Navier-Stokes) which is quite an impressive feat. However, to my knowledge nobody has created a worm model that fully captures actual worm behavior fully.
I think beyond mouse we will have to explore ways to upscale lower resolution images to higher resolution images using deep learning something like this https://arxiv.org/abs/2103.04989 or this https://www.youtube.com/watch?v=WCAF3PNEc_c&t . But for that first we need sufficient high resolution data to train the algos.