I'm fascinated that brain is compartmentalized enough, despite all cells being connected to each other, so a "local circuitry" can be identified for a specific action. I'd presume all cells would be firing more or less for any action. The existence of circuitry like structures makes it less organic and less mysterious for some reason. :)
Now I wonder if artificial neural networks exhibit the same circuitry-like structures.
They do. For example, training transformers on addition creates circuits that perform modular addition. See "A Mechanistic Interpretability Analysis of Grokking"
Having learned about computers before learning about the brain I assumed that the “grey matter” in our head was a general purpose substrate. However the brain is very much a set of heterogeneous processors with well defined purposes. We have evolved for a limited number of purposes, sadly.
Not all cells connect to each other remember and there are many different cell types with different functions which form functions specific circuits. Cell functionality goes far beyond the simple excitation-inhibition model
How many degrees of separation are there between cells? Many of them are pretty distant from each other even if they are technically part of the same overall network, we also have to consider the directionality of neuron signalling
Apparently, scientists a few years ago figured out how a mouse's brain encodes direction. They were able to predict the shape the encoding would take, and found it in the mouse's brain. They then were able to use their findings to determine which way the mouse was facing.
>We used a chemogenetic approach where we targeted the inhibitory designer receptor hM4Di selectively to SC-projecting RSP neurons by injecting AAVretro-cre in the SC and Cre-dependent hM4Di in the RSP (Fig. 2a and Extended Data Fig. 3)
My understanding of this is they used a virus to cause specific neurons to express the inhibitory hM4Di receptor which will modify the behavior of the desired target circuits with exogenous pharmacological manipulation when they introduce the agonist clozapine-N-oxide (CNO)
> virus to force specific neurons to express the inhibitory hM4Di receptor which will modify the behavior of the desired target circuits with exogenous pharmacological manipulation when they introduce the agonist clozapine-N-oxide (CNO)
Using these viruses is currently pretty common for things like optogenetics and tract tracing, it's some pretty neat stuff because they can be so selective in outlining specific circuits
IMO optogenetics can be a little more difficult to manage depending on the specific circuit you're studying and the recording method. Remember, you need to deliver the light to the actual neuron being manipulated and with in vivo models you're kinda limited. There are some recording probes with light guides built in but again, this method is limited and doesn't work super well for longer term recording projects
Personally I like pharmacological manipulation, but I'm probably a bit biased in that regard :)
yeah it seems much less practical, but it is pretty crazy from the standpoint of how it can be controlled - turn off and in by a literal switch.
There’s a dystopian black mirror future in here somewhere where people can take be administered viral vectors programmed with unique, proprietary methods of activation to stimulate a response - altering behavior, mood, hormones, etc… probably in a subscription model serviced by Comcast
I think that would be an incredibly difficult task to accomplish. Currently these viral methods require a precise injection to a very specific part of the brain depending on what you're trying to modify or study, it's not like you can just give someone a pill that does this stuff. In optogenetics, light needs to be delivered inside the brain or to the neurons a brain slice or culture.
That's on top of the fact that many behaviors are highly complex involving numerous circuits.
You can already give someone drugs or take a drug yourself to modify behaviors and we get greater specificity with newer generation drugs as we learn more about the affinity of ligands with subunits expressed in various circuits naturally.
Modification of behavior and thought processes is already a reality and it is in the field of psychological and behavioral neuropharmacology where we study these methods to help people with mental illnesses and neurological abnormalities
another technique sometimes used is using Transcranial Magnetic Stimulation to create temporary "virtual lesions" to study how knocking out a particular region affects task peformance
Now I wonder if artificial neural networks exhibit the same circuitry-like structures.