Metasynapses?

Science
1

I wonder if theres any biological circuit that could implement a network that outputs weights to operate anothee network.

I think it falls under the realm of dynamic gating.

I know there are the so gain field neurons that sorta implement this in a similar way to HTM but I believe it would have a problem with combinatorial explosion if you try to do it with arbitrary activity.

I wonder if there are circuits with synapses that connect directly to dentrite branches or something similar anywhere in the brain.

2 Likes
2

Attention has to work somehow.

I was fancying these days on an implementation variant (with no regards whether it makes sense biologically)

Gating methods usually either route input through one or a couple small “specialist” sub-networks (out of dozens-hundreds of sub-networks) at every layer, or they work directly on a large network where the router basically selects only a few parameters that will be active for the new input.

This one is a bit different, instead of picking 1 out of N small networks or shrinking a large network, it builds a small network dynamically with weights stored in a very large associative database and after learning step it saves them back in the memory.

Thus the associative memory, instead of storing answers it stores the tiny bits to dynamically build up a “learning block” trained to provide answers.
The advantage is each incarnation of the learning block can be a simple, one hidden layer MLP or even a single ff matrix and can be trained “classically” with whatever means (e.g. backpropagation). After each training step the updated weights are flushed back in their respective associative memory locations.

2 Likes
3

You may want to to take a look at the neuromodulatory system. I think it’s behaving in a way that is similar to what you are looking for.

@Bitking can probably correct me on this if I misremember any of the details; but IIRC, the neuromodulators are produced by cells in the basal ganglia and are released to specific areas in the neocortex. There are at least four major neuromodulatory systems (noradrenergic, serotonergic, dopaminergic, and cholinergic) that are triggered by internal and external factors such as hunger, thirst, danger, novelty, etc. Their immediate effect is to either block or enhance the uptake of certain neurotransmitters at synapses. Cumulatively, they introduce a kind of bias into the cortical network that acts to enhance or suppress the activity of a population of neurons.

2 Likes
4

Asking GPT-4:
What are the targets of subcortical neuromodulators in the cortex

ChatGPT

Subcortical neuromodulators are chemicals produced in certain regions of the brain that regulate neuronal activity and can alter the way the brain processes information. Common neuromodulators include dopamine, serotonin, acetylcholine, and norepinephrine. These chemicals are released by subcortical structures such as the ventral tegmental area (VTA), the substantia nigra, the basal forebrain, the raphe nuclei, and the locus coeruleus. They modulate cortical activity through diverse receptors distributed across different cortical layers and cell types.

Here are some targets and influences of subcortical neuromodulators in the cortex:

  1. Dopamine: Produced mainly in the VTA and the substantia nigra, dopamine targets several areas of the cortex. Dopamine is most famously associated with the reward system of the brain and plays a critical role in the prefrontal cortex, where it influences working memory, decision-making, and cognitive control.
  2. Serotonin: Originating from the raphe nuclei, serotonin modulates various cortical areas, influencing mood, anxiety, and cognitive functions such as learning and memory. One of the primary cortical targets of serotonin is the prefrontal cortex.
  3. Acetylcholine: Produced mainly in the basal forebrain, acetylcholine modulates cortical excitability and influences attention, learning, and memory. Cholinergic neurons are particularly influential in the primary sensory cortices and the prefrontal cortex.
  4. Norepinephrine: Produced in the locus coeruleus, norepinephrine has a wide range of effects on cortical areas, including attention, arousal, sleep-wake cycle, learning, and memory. This neuromodulator affects multiple cortical regions, with notable effects in the prefrontal cortex.
  5. GABA (gamma-Aminobutyric acid): While not traditionally classified as a neuromodulator, GABA is worth mentioning due to its role as the primary inhibitory neurotransmitter in the brain. GABAergic neurons in the basal ganglia and other subcortical areas project to cortical regions, where they play a key role in modulating cortical excitability and information processing.
2 Likes
5

Can you differentiate between direct physical connections and blood born neurotransmitters?

ChatGPT

Sure, I can clarify the difference between these two mechanisms of neural signaling.

  1. Direct Physical Connections: Neurons in the brain and the rest of the nervous system communicate via direct physical connections known as synapses. This process involves the transmission of electrical signals along the axon of the neuron. When these electrical signals reach the end of the neuron (the synaptic terminal), they cause the release of chemicals called neurotransmitters. These neurotransmitters cross the tiny gap (the synaptic cleft) between neurons and bind to specific receptors on the adjacent neuron (the postsynaptic neuron). This process can either excite or inhibit the postsynaptic neuron, leading to an electrical response. This is the primary mechanism of rapid, precise signaling in the nervous system.
  2. Blood-Borne Neurotransmitters or Neuromodulators: In contrast, some signaling molecules in the brain are released into the bloodstream, where they can travel to different parts of the body, including the brain. These molecules can affect neural activity by binding to receptors on neurons. For example, hormones released from the pituitary gland into the bloodstream can travel to the brain and influence neural activity. These types of signals tend to be slower and less precise than synaptic signaling, but they can have a broad impact on the brain and body.

It’s worth noting that some molecules, such as dopamine, can function both as neurotransmitters (in synaptic signaling) and as hormones (when released into the blood).

The mode of signaling can greatly impact the effect of a molecule. For example, a molecule acting through synaptic signaling might have a very precise, localized effect, while the same molecule acting as a hormone could have a widespread, generalized effect. This is one reason why the nervous and endocrine systems can regulate a wide variety of bodily functions and behaviors.

Can you list the blood born neurotransmitters and neuromodulators?

ChatGPT

Neurotransmitters and neuromodulators can be released into the bloodstream, functioning in a similar way to hormones. They are produced by various organs and tissues in the body, including the brain and gut, and can affect a wide range of bodily functions. The following are examples of neurotransmitters and neuromodulators that can be found in the blood:

  1. Serotonin: More than 90% of the body’s serotonin is produced in the gut. It plays roles in the regulation of mood, appetite, and sleep, among other things.
  2. Dopamine: While dopamine acts primarily as a neurotransmitter in the brain, it can also enter the bloodstream where it affects heart rate and blood vessel function.
  3. Epinephrine (Adrenaline) and Norepinephrine (Noradrenaline): These are both neurotransmitters and hormones. They are produced in the adrenal glands and certain types of neurons and play critical roles in the body’s stress response.
  4. Acetylcholine: Acetylcholine is primarily a neurotransmitter, but it also functions as a neuromodulator. It’s involved in muscle contraction, memory, and other functions.
  5. Glutamate and GABA: These are the main excitatory and inhibitory neurotransmitters in the brain, respectively. They can also be found in the bloodstream, but their function there is not as well understood.
  6. Oxytocin: While oxytocin is known as a hormone, it’s also considered a neuromodulator. It’s involved in social bonding, sexual reproduction, childbirth, and the period after childbirth.
  7. Endorphins: These are neurotransmitters that also function as hormones. They’re often referred to as natural painkillers because they can block pain signals.
  8. Cortisol: It is primarily a hormone, released in response to stress and low blood glucose levels. It also has neuromodulatory effects on the brain.

This is not an exhaustive list, and the exact role of these chemicals can vary depending on where in the body they’re acting, the specific receptors they’re binding to, and other factors.

Same question, but can you provide an exhaustive list?

ChatGPT

Creating a comprehensive list of all neurotransmitters and neuromodulators that can be found in the blood can be challenging. The classification of molecules as neurotransmitters or hormones (which would make them blood-borne) is sometimes blurry, and research is continually discovering new signaling molecules and new roles for known molecules. However, I’ll provide a more extensive list:

  1. Catecholamines: This group includes dopamine, norepinephrine, and epinephrine. These neurotransmitters are involved in attention, focus, and the fight-or-flight response.
  2. Serotonin: This neurotransmitter is involved in mood regulation, and its modulation is the target of many antidepressant drugs.
  3. Acetylcholine: This neurotransmitter is critical for muscle movement and cognitive function.
  4. Histamine: This molecule acts as a neurotransmitter in the brain but also has many roles outside the nervous system, including in the immune response and regulation of stomach acid.
  5. Glutamate and GABA: These are the primary excitatory and inhibitory neurotransmitters in the brain, respectively. They are also found in the blood, though their roles there are less well understood.
  6. Glycine: This neurotransmitter is predominantly inhibitory and is crucial for a variety of physiological processes.
  7. Purines: This group includes adenosine and ATP, which act as neurotransmitters in some contexts.
  8. Peptides: These include substances like oxytocin, vasopressin, endorphins, and many others. They often function as hormones but can also have neuromodulatory effects.
  9. Endocannabinoids: These are naturally occurring compounds that influence the cannabinoid receptors in the brain and other tissues, impacting a variety of physiological processes.
  10. Nitric Oxide (NO): This molecule is unique because it’s a gas. It functions in various contexts, including as a neurotransmitter.
  11. Trace amines: This group includes compounds like tyramine, tryptamine, and phenylethylamine, which are structurally related to the classical neurotransmitters and can have neuromodulatory effects.
  12. Steroid hormones: This group includes glucocorticoids like cortisol and mineralocorticoids like aldosterone, as well as sex hormones like testosterone and estrogen. These hormones can pass through the cell membrane and directly influence gene expression, and they also have significant neuromodulatory effects.

Remember, the roles of these substances can be very different depending on where they’re acting and the specific receptors they’re binding to. This is not a comprehensive list, as there are potentially many other substances that act as neurotransmitters or neuromodulators and can be found in the blood.

1 Like
6

More on this from an older post:

7

I think that the OP is thinking like a digital computer engineer where a signal path is gated from one area to another. I don’t think of processing in the brain as working like that. The connections are always there, in fact they have to be there if an association will ever be formed.

I see the basic function of a map/region/area as being pattern completion. This is completely compatibly with HTM theory.

As sensations (or probing patterns in the counter-flowing direction) impinge on a area, the area works to recall the best possible match. A key feature is that every area after the primary sensory cortex has more than one source of input fibers, so the recalled pattern is the best match of a combination of features.

This is the essence of the critical “sensor fusion” function in the brain.

It is also how you search through your stored memories in response to your forebrain driving the speaking or thinking process.

If this interest you I could start a separate thread.

8

I was actually thinking more like a graphics programmer where you need to transform data using a reference frame before you can render it.

Its just a guess but I think thats nescessary for the sort of what-where encodings that happen in the visual pathways and I have less confidence in this but I think it would be useful for language and speech too.

Just as an example…

MNIST is a very easy dataset, thats because all digits are perfectly centered.

but if those digits were to move randomly in a larger image, the problem becomes harder, still you can just compute a centroid of the pixels and “unshift” the digit.

but what if the pixels are randomly permuted and you dont know the order, how do you unshift them?

my thought is that we need three networks.

One: a basic spatialpooler-like encoder that turns the input into a sdr.

Two: a “what” network that tries to classify which digit it is, regardless of where it is.

Three: a “where” pathway that tries to learn a representation of the location of an object, regardless of what it is.

putting what and where together allows you to reconstruct the input but either one alone is insuficient.

the way I think this could work is if the where pathway was used to generate a “shifting” matrix to stablize the input so that the what pathway can recognize it.

1 Like
9

I am not aware of any structure like that in the cortex.

10

Chiming in a few thoughts here…

  • Basically any neuron synapses triggered below activation threshold can “amplify” any signal from the other synapses.
    e.g. a neuron triggered by a duck, could get a continuous input from both higher levels expectations “a duck is likely near the shore” and visual cues.

  • about MNIST performance - I’m pretty sure fovea vision makes image recognition pretty insensitive to position and, to a lesser degree, to its size. Instead of treating vision as a passive sensory organ, I would consider it as an agent, capable to actively move into, search for specified patterns, focusing on parts of the available scene and abstract out other parts of it. It is never a one frame encodes-it-all kind of game computer/ai vision is trying to perform.

Edit:

  • spatial correlations for randomly connected pixels can be inferred, at least partially by counting how often two arbitrary pixels (a pair of pixels) read the same value. The closer two pixels are together the more likely they will see the same “color”. So with a reasonable amount of “looking around” one can build up a distance metric between pairs of pixels or even regions.
d(B,C) ~= d(A,B) < d(A,C)

means B is somewhere in between A and C

However there are reasons to believe that visual systems somehow “mechanically” encode position.

1 Like
11

its true that eyes can do that but you know, sometimes I simply dont bother looking directly at stuff, for example, I’m cooking and want to pick a spoon and instead of looking for it I just see it in the pheriphery of my vision and pick it up accuratelly. also killing flies and mosquitos, they are too fast for my eyes to follow yet I can still track and catch them. also watching movies in another language with legends, I can read the captions with my lower pheripheral vision just fine.

to do this kind of stuff, I think its nescessary to have a dynamic routing and some form remapping of inputs based on object locations.

1 Like
12

my takeaway is that I’m pretty sure I’m able to recognize a spoon anywhere in my visual field but I’m also farily sure I havent lived long enough to have seen a spoon in every possible location.

1 Like
13

Gaze tracking studies have shown that much of your visual parsing is below the level of awareness.

You might be surprised at how much you think of as “peripheral vision” is the result of a quick visits from your foveal vision.

Next time you are bored, try to perceive your saccades shifting to take in facial features. It is extremely fast and as far as consciousness is concerned, automatic. Your eyes see 1.7 degree snapshots, you see a complete face.

There are many experiments performed with the eyes fixated on a spot and you certainly do have perception of the areas away from the fovea.

When reading text outside the fovea vision area, your brain relies more on contextual clues, word shapes, and patterns to interpret the text. While you may not perceive the text as sharply as when it’s in your foveal vision, you can still gather meaning from it, but I suspect that you are taking quick trips to the text to read it and get this context.

1 Like
14

actually I might not be exactly neurotypical in this aspect.

I have visual snow and kinda literally perceive the world at 11-15 Hz, that is to say the illusion of temporal continuity is broken to me.

I’m also awfully aware of each and every sacade, even microsacades so I’m not fond of darting my eyes around as it makes me a little dizzy.

2 Likes
15

I find that interesting in itself, looking forward to hear about your objections to our models based on that particular perception.

That being said:
Yes, It’s probably not ‘convolutions everywhere’. Although I do think you’ve lived long enough for the ‘edge-detection’ convolutions to at least kick in everywhere. Based on that, there may be further convolutions, or there may not be, or there are both… I believe there are still convolution-like fast-paths to higher abstraction levels… towards the still-possibly full-convolution (and fully subconscious) ‘spoon’ recognition.

Since you doing cooking and grabbing that spoon aside, or you doing long, conscious, peripheral-vision experiments is not 100% proof that there is no such high level convolutions, though:

  • grabbing the spoon while other parts of your brain are probably aware that you dropped it before in that general direction is probably enough to tell that “unidentified strange edge #1138 on the left” and “should be a spoon on the left” could subconsciously-enough match to “spoon there”.
  • Having conscious processing and loops to forebrain, in my other example, is a clue that various other brain loops can interfere with the process of pattern recognition, including high forebrain logic and inference rules and all kinds of stuff… able to somehow resolve to “spoon” in the end, even when the convolution-spoon failed to match directly. But as you can probably tell, this loop-to-forebrain has a cost and a noticeable longer delay.

As an aside, “lived long enough for xyz” should hint that it is an intrinsically plastic and experience-based process. If you’d been totally blind, well… you’d still possibly have lived long enough for catching flies or mosquitoes by sound, and it doesn’t really tell how the eye&brain would have learned in the general case to detect flies, see what I mean ?

1 Like
16

Neurotubule Cytoskeleton (just a marker for future reference).

1 Like
17

nah, at arms reach distance, spoons are fully recognizable for me up to around 70° of visual field after that they are more like “unindentified blurry blob with suspicously concave metalic shine next to unindentified edge”

1 Like
18

on a side note, I diont quite understand why people treat pheripheral vision as if it was “semantically blind” its true that its not quite as sharp as central vision but for me, it feels just like regular foveal vision but with less definition.

just like the skin on fingers has better touch resolution than your knees but it doesnt mean you cant feel anything on them.

I think that the idea about us only seeing with central vision is more of an exageration otherwize there woudnt be that talk about overt vs covert attention.

19

Central vision loss is considered a serious condition. Depending on the severity, many “normal” activities such as reading are impaired. In advanced cases it is considered legal blindness and can cause you to be barred from driving.

20

of course since resolution drops exponnetially with excentricity but that doest mean peripheral vision is blind or processed in differently than central vision. I suppose losing central vision is a problem only because of lower resolution and because we tend to sacade into the direction we attend to, effectively putting a blind spot where we are trying to see.

still, I am unable to believe the processing of sight is done differently, I think receptive fields are just larger and sparser and diminated by the magnocelular stream.