I read a few Numenta papers, but I have a very few knowledge of neuroscience and I don’t know where to start (I have read “On intelligence” book already).
Particulary, I would like to know how connections of neurons in the brain were formed. I understand that the most of them were “prewired” (I’m not sure what term is correct) during the development of the brain and part of them were created during learning, but it’s not clear how many connections were “prewired”. Is there some kind of “random initialization of the weights” like in neural networks involved?
I’ll be glad if someone will help me to understand this topic or point to some books or articles where I can find the answers.
You are not alone - many people would like to know this in greater detail
Neuroscientist armed with microscopes and stain mapped out the major brain locations in the early 1900’s.
Since that time there have been countless studies connecting observed defects with lesions in these areas to guess at what functions each area might perform.
Sorting out the huge glop of white matter connecting these areas has been slow and uncertain. Many areas connect to more than one other area and the connections run both ways. It is very difficult to tell what is going on there.
Check out some of the pictures in this google search to get some idea of how nasty a task this is:
Google: human brain interconnections map (Then click “show as image”)
I think this link illustrates the problem particularly well:
That said - there has been some good work in this area. An example is the Papez circuit.
This was mapped out by putting a marker in one area and following the fibers to where the marker wicks up to. Repeat for the next area.
The current best effort is the Human Connectome Project. These folk are using various imaging technology to trace out these paths. They have a downloadable viewer and some great datasets to load into the viewer.
(Pro tip: installation is kind of fiddly. Read the documentation and work through the examples to the letter before trying to just click the buttons) I have spent many hours looking at this myself trying to sort out how these connections match up with what I know about the functions that are connected. I expect much more to come from this. One of the first exciting findings is that the number of functional areas is roughly twice as many as Brodmann discovered using the methods available to him.
BTW: This is what the major interconnections look like in 3D:
Moved from #htm-theory:neuroscience (because that is really the place for more advanced discussions on the topic) and changed the title to “Neurosciene newbie questions” so others might feel welcome to post their “uninformed” questions here.
If you want to zoom in on the neural hardware much has been mapped out. There is an ocean of books and papers on this and sorting through all this can be daunting. I offer this tutorial as a nice start on this area of study. When this material starts making sense you will be able to sort through google searches to answer other questions as your learning progresses.
This will get you through the basic neuroscience needed to understand how HTM/SDR model neurons work and how they work together to form columns. http://www.mrc.uidaho.edu/~rwells/techdocs/Cortical%20Neurons%20and%20Circuits.pdf
Hi BitKing
Isn’t the simple answer that the cells in the sense organ are prewired to those cells in the neo-cortex which receive their input? And each pair - the sensory detection cells which detect input and neo-cortex receiving cells which are dedicated to receiving that specific input from those particular cells - are joined by a single, dedicated, nerve fibre which must, therefore, have been a pre-wired connection since it is a hardwired connection dedicated to carrying spikes from one to the other but no others?
Lets say a yellow light is detected by 8 seperate groups of retinal cells. They each send a spike which activates 8 separate groups of cells on the neo-cortex. Those inputs reinforce and ‘clean up’ each other’s signals and in this way lead to the creation of 8 active cells. Those 8 active cells form the ‘ones’ in a single SDR. That SDR then goes on to create / encode further patterns of activation in the NNs and thus new SDRs.
So in response to @aetolicus’ question; the connections from the sense organ to the neo-cortex are pre-wired and the connections between neurons in the brain which process the input further up the heirarchy are created during processing as they learn the new temporal patterns of activity,
Here is a little experiment you can do to explore that idea.
Let’s be your 8 cells and get an idea of the job they are doing.
Poke 8 random holes in a piece of opaque paper, say 1/16 inch (1 mm if you are like that) size.
Lay that on a TV screen and turn down the sound. Make sure all you can see of the image is from the 8 holes.
Now you tell me what is happening on the screen from that information as the video progresses. Don’t forget to slide it here and there as your logical eye moves around.
Just for fun, do it with a movie you have seen before.
Somehow there is more to it than just learning a single pattern. That is the special sauce that makes this interesting.
Even if you focus on the 8 cells and poke 8 holes in the paper, the other neurons and retina cells are still active and the entire input stream is being processed to infer about the movie. The 8 holes will obviously have lot more than 8 cells processing its input. And the data from the 8 holes will be compared with the patterns that have similar data(from the 8 holes) as their subset using different mechanisms.
And how does this pertain to hardwired connections?
1:1 in what sense?
The data the cells are processing might be different, but voting and comparing also takes place at different scales and using high order representations, up the hierarchy.
And how does this relate to hardwired connections? Even if the connections are not genetically determined, once they are decided, i don’t think they change. Are you suggesting the influence of other brain areas in localising certain input patterns to certain regions?
I never meant to imply that connections change anywhere in the brain or sensory system.
The hardwiring is surly programmed in the genome.
I can’t see how anything I said could be taken to imply that any of this is not true.
I am contesting that conditioned behavior at the local sensory level yields to object recognition. I can’t draw a line from (what this or that neuron senses) to the (formation of grid cells oriented to the external surrounds) and not the critters (internal spatial reference). Likewise, an object can be recognized even if the position relative to the critter changes, or is even completely novel. I can even recognize an object that is similar to one I know but that I have never seen before.
This is all true even though all cells in the chain are hard-wired to connect to their targets in various neural sheets. The synapses vary connection strength to learn things (and form connections) but even those connections are relatively stable in the short term of navigating and parsing the world to the point of recognition. The fast dynamic connections are activation patterns formed on the neural sheets.
I believe that it takes mass action to extract object recognition using the WHAT and WHERE streams. The stuff being fed to these streams is local features. These features are given meaning at much higher levels.
Perhaps this is rephrasing what Bob said just above but I think that the processing at higher levels is of a different nature than I read from that comment. I have to add that I came to this comment after reading Bob’s post in the Universal Encoder area.
Aw, come on Bitking, that is a straw man argument.
I meant we take the 8 cells at position r(x1y1) on the retina that only detect yellow. And assume that their spikes arrive on V1 at v(x1y1). As I said, those inputs create an SDR with 8 ‘ones’, each ‘one’ a cell that is produced by, and only produced by, yellow at x1y1.
Hey, @aetolicus,
I’m a newbie too, so what I do is imagine that the input arrives in a column of cells in the neo-cortex. It arrives in a layer, say the layer contains 2040 cells and maybe 40 of them are active. I suppose that makes as SDR of 2000 non-active cells - usually shown as ‘0’s and 40 active ones - usually depicted as ‘ones’ or blue squares. The column has a ‘feed forward’ system which moves the ‘informational load’ of the SDR up to the next layer of cells.
To my mind the connection, the nerve fibre that transports the input from sense organ to V1, has to be both prewired and hardwired so the brain can be sure that the arrival of those spikes always signify yellow at x1y1. If not then the brain would never know their significance! But having got that activity in the HTM it can do what ever it wants [eg learns, adapts & changes the patterns of activity]. Tho’ whatever it does with the effect of the activation of those 8 input cells that effect has to continue to mean [eg play the role in the computational economy of] yellow at x1y1.
I agree. I can’t see how one can have any meaningful input unless the connection between the sensory organ and the input layer is hard wired. But aetolus’ question also implies that we need to consider what happens in layers other than the input layer. Once we accept that the input layer has cells dedicated to a specific meaning then why can’t we ‘assume’ that there are other layers also containing cells which are dedicated to, for instance, yellow at x1y1. Indeed it would be hard to have a swamp of non-dedicated cells, cells which are continually changing their significance, unless… [but that starts to get really complicated] … so one might just as well assume we have to have cells scattered through out the brain which only and always mean ‘yellow at x1y1’!
But I’m not sure if this is ‘official’ HTM theory. Bitking? Abshej? Help me?
I wouldn’t say that cells are assigned specific meanings. The input bit to neuron connection is hardwired, but what the input bit or neuron represents depends on the encoding, and level of abstraction. So in a way in a developed system you could say that say that a set of neurons fire when a certain feature is detected in the input, but these meanings that are encoder dependent and I find it troubling to establish those neuron-semantic relationships as permanent, because of the nature of HTM.
Plus the connections with the input space could increment of decrement depending on the presence or absence of on bits.
I am inclined toward to saying you are right in saying it the way you did, but I find it difficult.
{If we use the terminology, from the diagrams, of sheets of squares. Where each square = one cell. 0 = white, 1 = blue.}
I thought that - as in semantic folding - the position was = to the meaning or informational load. And, as Bitking was saying, only the synapses / weighting changes. Sure, if one goes up several layers [of neural sheets] then a blue square may have the meaning of ‘yellow @ x1y1 moving to the right’. But it would map back to those blue squares meaning either ‘yellow @ x1y1’ or ‘moving to the right’. If not, how are you ever going to know what you’re thinking about?
I agree… As I said, it’s just that it is difficult to accept that the meanings were genetically assigned to the neurons. It could be that it is a coincidence(while forming connections) that a particular neuron got connected to the bit representing a particular feature.
ah, yes, I see. I agree, its v hard to comprehend how the brain could have so many pre-wired connections. Perhaps this is why there is so much pruning in a baby’s brain.
I’m not a neuroscientist, but personally I have a hard time believing such specific connections as “yellow at x=1, y=1” could really be fit into our genetics. It would be much easier to accept that “something at somewhere” is what is encoded in our genetics (where the specific neurons for “something” and “somewhere” are different for every individual). The fact that “something” is “yellow” and “somewhere” is “x=1, y=1” in a particular state can be learned via temporal associations.
