Neuronal signaling: All or none?

Most computational models build off this ‘all or none’ assumption, and they seem to do a decent job in representing some phenomena. But why would it have to be this way? What is so special about this transient fluctuation in membrane potential, and what about it allows images and events to enter our awareness? Is it the opening of voltage-gated ion channels? Release of neurotransmitter? Rapid movement of ions? Do action potentials that originate at the hillock have different effects than those initiated in the dendrites?

I cannot wait for the day when we have the technology to begin to address some of these questions. And I think we’re close. With techniques like 2-photon imaging and calcium uncaging, we could theoretically “excite” neuronal boutons such that neurotransmitters are released (via Ca2+ binding) without an action potential ever taking place in the cell. Conversely, we could block neurotransmitter release and electrically stimulate cells. Granted, information is almost certainly held within vast networks of cells, so experiments like these on the single-cell level are futile. But enormous progress is being made in identifying neuronal ensembles active during a particular behavior or memory (for example, see this week's J Neurosci article from Dombeck et al). And if we can identify the neural correlates of a particular memory, we might be able to apply techniques similar to those mentioned above to this population.

But the question still remains: what the hell is so special about action potentials? And can postsynaptic potentials that fail to reach threshold still convey information at the level of awareness?

I also need to mention a wonderful review article by Alcino Silva and colleagues in a recent issue of Science, titled Molecular and Cellular Approaches to Memory Allocation in Neural Circuits. In particular, there was one assertion that thoroughly confused me.

Can proximity of strengthened synapses affect co-recall?

How does the proximity of synapses (on the same neuron) have anything to do with the likelihood of co-recall??? Based on the binary action potential assumption above, an information signal is based on whether a neuron fires or not. And if a neuron fires, it should have the same effect on all it's afferent connections. Far as I know, activated spines do not send out retrograde signals that potentiate surrounding synapses on an immediate time scale. So, what am I missing???

Seriously, I’m asking you. YOU, person reading this right now. You’ve read this far, so you must be interested. The only thing I can come up with is that events occurring in close temporal proximity lead to increased probability of potentiating synapses on the same neuron. And if there is large overlap of neurons that make up the networks representing events A and B, then the probability of activating network A will increase the probability of activating network B. But in this scenario the proximity of these synapses should not matter, so long as the synapses reside on the same neuron. WTF???

One last question: could the originating location of an EPSP be a source of information?

Comments?

Synaptic tagging?

Without reading the review, could they not be talking about events after the action potential has fired? When the first memory is triggered by activation of a certain set of synapses, molecular events at these synapses can alter them in some way to modify the probability they will fire next time. If the specificity of these molecular events is not perfect, synapses in close proximity to the activated set, may also get tagged or modified, thus changing their firing probability, which in turn will affect the ability to recall the associated memory. All the synapses are still firing all-or-none (don't worry, the universe didn't just turn upside down!) but events after the initial action potential can influence a local region of dendrite, which will include the synapses of the second memory. This paper from the Svoboda lab talks about how molecular tagging events are sometimes not localized to just the activated synapse but can spread a little way along the dendrite. This may not be what you were getting at Jeremy, but it's the only explanation I could come up with! -Clare

Hmmmmm, still don't see it

Bloody hell. Just responded via word press, but lost it due to a dang error page. Let's try again.

Yeah, I see where you're going with this, and I think it's an excellent explanation of why two events in close temporal proximity (A and B) would be stored in neighboring synapses. But I fail to see how this would affect co-recall of these two memories. In my mind, co-recall is the process where retrieval of memory A increases the probability of activation of memory B, and this takes place on the scale of sub-seconds. However, the passive diffusion of agents that Svoboda discusses takes place on order of seconds, yes? Now, I could see how retrieval of memory A could facilitate retrieval of memory B at some point in the future, but co-recall? Maybe I'm misinterpreting their definition of co-recall (or your response)?

But even if the diffusion and influence of these potentiating agents were instantaneous, would that really matter? In this case, activation of memory A might make it easier to independently recall memory B, but it wouldn't activate memory B, and thus no co-recall. I suppose I could find some scenario using reverberating loops or something where this might work, but I don't think that's what the authors had in mind.

Something ain't right. Either I'm totally missing something, my base assumptions of how the brain operates are fundamentally flawed, or the review is mistaken. I suspect it's the first.

Jeremy Biane

Comment on Qualitative Information

Bradley Monakhos 23:16, 9 November 2009 (UTC)

Fig 1. There are about 100 million photoreceptors in the human retina and only 1 million ganglion cells; on the average, 100 photoreceptors must be connected to 1 ganglion cell.

Everything in our environment comes to us through waves of energy. When a wave contacts a surface, two piece of information can be coded temporally - Frequency and Amplitude. Coding the frequency and amplitude is enough to transmit qualitative information about a stimulus. In the visual system for example, qualitative information must somehow be transmitted through modulation of neuron firing patterns because of the quick convergence of stimuli information passed from photoreceptors to ganglion cells (see Fig 1). Although, spatial information is stored retinotopically, qualitative information must some how be conveyed by information stored in the frequency of neuronal firing patterns. This is the same for light, sound, touch, pain, etc.

--- Dees Noodle Need Sauce I know, you need a source right. Here's a little snip from a researcher at the RIKEN Institute. If you don't know what RIKEN is, it's the MIT of Japan. In fact, many of the faculty at MIT and RIKEN have joint appointments at both institutes. Be sure to check out the RIKEN-MIT Center for Neural Circuit Genetics.

Enter Toshihiko Hosoya:

“It has been believed that a single neuron usually transmits a single piece of information. However, doesn’t that seem wasteful?” says Hosoya. “Information will be transmitted more efficiently if multiple pieces of information are carried in one signal. This also forms the fundamental basis of information processing theory.”

Therefore Hosoya decided to investigate extensively what information is transmitted by the retinal neurons. He used the neuron that responds to OFF, which ‘fires’ (generates an electrical signal) when it becomes dark. First, he repeatedly stimulated the retina using an image with changing dark areas, and examined the timing of the firing of the neuron that responds to OFF. It used to be thought that the responses of the retinal neurons involved a wide temporal fluctuation, with the timing of the firing differing between events, even in response to the same stimulation. In recent years, however, it has become evident that firing occurs with high reproducibility in terms of time in response to the same stimulation. Hosoya showed experimentally that the firing events in response to the same stimulation occurred with only a small fluctuation of several thousandths of a second. “This is quite an intriguing feature,” says Hosoya. “Because the retinal neurons fire with high reproducibility in response to the same stimulation, it has become easy to analyze what kind of information is conveyed when the neuron fires.”

Fig 2. Responses of retinal neurons. Firing (green points) of OFF-responding neurons in the salamander retina on stimulation by the brightness change shown by the gray line. This represents data for a single neuron receiving the same brightness change for 21 iterations. Firings caused by each iteration are shown in a row; cases of three firings are colored. The intervals between the first and second firings and between the second and third firings are found to be highly reproducible between all iterations.

The retina neurons that respond to OFF fire more frequently as the image input becomes darker and darker. “It has long been known that the firing frequency changes depending on ‘how dark it has become,’ or on the amplitude of the intensity change. Extensive examination shows, however, that both the firing frequency and the time interval are quite accurate (Fig. 3). This seems to carry some information.”

Hosoya theorizes as follows. “The neurons may transmit information on ‘how it has become dark’ as well as on ‘how dark it has become’.” It has also been demonstrated that accurate firing patterns are transmitted from the retina to the brain through the optic nerve. Upsetting conventional common sense, it may be shown that a single neuron transmits multiple pieces of information by using an accurate firing pattern.

Commnet on 'Comment on Qualitative Information'

Ok, good foundational argument for your rate-encoding system. The conclusion of Hosoya seem quite intuitive to me, so I don't know what they're talking about when they say 'upsetting conventional common sense.' (Maybe I possess uncommon common sense?) It's easy to see how onset of firing encodes when a stimulus is present, while firing rate encodes amplitude. But I find it harder to relate this to internal circuits that code for something like a memory. No, wait, maybe I can.

Ok, in 'my' system where each neuron in the circuit encodes a particular aspect of the memory, firing rate could simply encode amplitude of that aspect. For instance, with our black neuron (see discussion below), rate of firing would encode the amplitude of black of the object. This scenario would be interesting in that you'd have neurons belonging to the same memory circuit firing at vastly different rates, each encoding for the amplitude of their respective representation. Ok, so there. Now we have a testable hypothesis. Go prove me wrong :)

Jeremy Biane

Wordpress Comments

• admin Says:
  Nov 05 2009 8:22 pm
  

  Neuronal signaling

• Bradley Monakhos Says:
  Nov 05 2009 8:28 pm
  

  Temporal summation at the axon hillock? Could it be that neurons that make synaptic connections in close spatial proximity evoke similar dendritic potentials on downstream neurons. Voltage that propagate as graded membrane responses to the synapses with a second-order neuron decays exponentially with the distance from the synapse, these potentials may in turn be reaching the hillock with related information (magnitude and/or frequency).

• Jeremy Biane Says:
  Nov 05 2009 10:06 pm
  

  The above is in response to the last question I threw out there at the end of the article, yes? ("could the originating location of an EPSP be a source of information?")

  If so, then I agree that the influence of an EPSP on the potential at the hillock can vary with respect to the origin of location and the frequency of stimulation. So, yes, I think that's the correct answer. But I also think it's the boring answer, as it again assumes that it's the AP that is really carrying the info.

  If your comment was in reference to the whole co-recall debacle, then I need you to elaborate a little more, because I can't quite see the connection. Unless...you're suggesting something totally novel. That is, memories aren't stored in ensembles of neurons per se, but in a combination of ensembles and firing rate. And since the influence of EPSP can depend on the location of the synapse, afferent cells with neighboring synapses should elicit similar effects on the postsynaptic cell, and maybe even several cells downstream. This way, you have an overlap of neurons and firing rate for encoding memories A and B. Anything remotely close to what you were thinking???

  Along a similar line of thinking, maybe neighboring synapses don't ensure similar firing rates, but a similar "location" of activation relative to local EEG cycles (fire at the same spot along the theta waves). Oooh, now I'm really stretching it. And could I be any more vague?

• Brad Says:
  Nov 05 2009 11:05 pm
  

  I can interperate your respose as, yes, that's what I was thinking. Do you remeber our convo a few years back when I was forewarding the idea that thoughts are represented as rhythems. These rhythems must somehow flow through a timecours, actively pinging the necessary sensory representations external stimuli. It seems to me that objects experinced visually in close temporal proximity... wait, how do they know which neurons are responsible for which thoughts?

• Jeremy Biane Says:
  Nov 06 2009 2:20 am
  

  Yeah, I remember our conversation from way back. I thought you were way off in thinking that objects/events were coded by firing rate, and not by specific networks. And while I still think a particular ensemble is responsible for, say, the concept of an apple, I'm beginning to warm to the idea that firing rate plays an additional role.

  "how do they know which neurons are responsible for which thoughts?" Several techniques are being developed for this sort of thing. One is imaging a population of neurons with calcium dyes, the idea being that when the neuron is depolarized, Ca2+ enters the cell, binds to the dye, and causes detectable flourescence. This can actually be done in vivo, but is very difficult.

  More "traditional" forms include tagging neurons that were active during a particular time window (remember that minor prop project I proposed, and how I wanted to specifically deactivate neurons that were engaged during fear conditioniong? Remember my strategy with the IEG promoter controlling the inhibitory receptor, and this whole construct being repressed by dietary doxycycline, thus allowing me to limit expression of my construct to a time window of ~one day?). All of these techniques usually rely on linking a reporter gene to an immediate-early gene promoter like c-fos, zif, or arc. That optogenic paper I wrote about used this technique, except they used the IEG promoter to drive expression of ChR2 instead of some visibly detectable gene.

  Then there's multi-unit recordings. Throw 100 electrodes in the brain, and see which events correlate to spiking behavior. You're only sampling a small number of the overall cells in this case, but you might be able to identify a few cells that are strongly correlated to some event or behavior.

  There are probably a few others I'm missing. Maybe they'll come to me in sleep.

• Brad Says:
  Nov 06 2009 4:11 am
  

  "objects/events were coded by firing rate, and not by specific networks." 

  Well, it must be both. As I'm thinking about this, I am trying to put together an example of how it would work - Just for a reference let's use a situation where are sitting at a bus stop and a black Jag drives by bumpin Citizen Cope, then 30 seconds later a dalmation with 3 legs starts barking at you. 20 min goes by and a pink caddy rolls past playin Jay-Z. The next day, you see a dalmation, and it's no trouble at all to recall the song that was playing in the Jag, but not the cadi. Better yet, let's say a month later you were just thinking about dogs by your own free will and Cope pops into your head.

  - that's an example we can continue with; moving on... 

  
  We know from TMS studies that when an area of the brain responsible for encoding/decoding a particular sense perception (ie color) gets knocked out, we can no longer recall colors in previously stored memories, even though the "data" is still there. Thus, the "network" for a particular memory is at least as dispersed as the specialized sensory areas of the brain, which are all over the place. But the sense areas are relatively small compared to the vast array of stimuli that need to make use of the "black" neurons (insert clever racial punchline here). Thus, if LTP is driving associative learning, it's probably happening at least a little bit upstream from the sense-encoding neural populations, because of the bottleneck that occurs right outside these areas. So how does the Black Jaguar memory go in, grab the color info it needs and come back out without "black" flowing down the, say, bowling ball, neural representation? If it was simply a network thing, then you might expect the i/o to produce far more unrelated thoughts than what we observe. It must somehow rely somewhat on the fact that the Black Jag neurons are more active at this time, and in order to get back to them on the color issue, it makes sense to me that the black neurons start firing in rhythem with the active neurons; else disturb the peace among the resting neurons that also map to black (albeit, with their own unique firing keycode). 

  "I also think it’s the boring answer, as it again assumes that it’s the AP that is really carrying the info."

  That's like, your opinion man

• Jerms Says:
  Nov 06 2009 12:32 pm
  

  First off, that TMS study - really? That's so cool!

  Next.

  True, the "black" neurons are going to be included in all sorts of networks - bowling balls, ants, jags, tacos...

  But why would all of these networks become activated simply by firing of this minute part of the network? I doubt that 5 black neurons firing would be enough to activate a downstream neuron, UNLESS that downstream target is simultaneously excited from additional input.

  For example, you have your two networks: bowling ball and jaguar. Three-legged dalmatian comes hobbling by. Because memory of dalmatian and jaguar were laid down in overlapping populations of neurons, network of black jag blasting distorted Cope via stock speakers is jogged. Black neurons are activated. Black jag comes to mind because majority of network is active. bowling ball does not come to mind because only a small part of that network is active. HOWEVER, because that network is somewhat active, if asked to think of something random, you'd be more likely to say bowling ball than, say, hamstring (unless you've had a strained hamstring for the last 3 months that's prevented you from playing any sports and is the bane of your existence).

  Also, the black neurons may preferentially activate the jag network because these downstream neurons are already excited by the dalmatian siting. Point being, coherence of firing rate is unnecessary in such a scenario.

  But, if your interpretation were correct, why would we need the rest of the network to fire at all. Wouldn't just firing the black neurons at the "black jag" firing rate code for black jag?

• Bradley Monakhos Says:
  Nov 06 2009 3:00 pm
  

  "But, if your interpretation were correct, why would we need the rest of the network to fire at all. Wouldn’t just firing the black neurons at the “black jag” firing rate code for black jag?"

  Super idea! How efficient.

  EFFICIENCY important for a system that processes 10 million bits/s through one retina alone.

  The other thing I don't understand about the LTP Network hypothesis is that it seems like it would be sooo energetically inefficient, slow, and probably would be wasting huge amounts of space. We know that non of these are the case.

  Say your watching cartoons a blue dalmatian comes onto the screen. The neural network encoding for the dalmatian now has to grow some arms and extend them on over to blue and shake it's hand. Then a red dalmatian, and then a pink, and then there is a scene with all types of animals in novel colors is on screen, all of which not only need to find their respective new color neurons, then need to synapse on it in close proximity. BS.

• Jeremy Biane Says:
  Nov 06 2009 10:11 pm
  

  But how would a system like yours ever be trained? In the LTP system, it's not like novel connections between neurons are being made. All these neurons are already linked up to one another, just the connections are strengthened/weakened with experience. So, it's not like a blue dalmatian would require new connections. All the base elements (and connections) are already there. But with your single neuron system, you'd have to create a new rate/neuron combination for every object. And it seems to me like the representation of a blue dalmatian would have to be plucked out of thin air.

  And maybe it's good that plasticity in the brain is a laborious process. It would be quite detrimental if our brains were in constant flux, ready to remap their contents for every microenvironment we encounter.

  So, how do you envision your system integrating new knowledge into an existing structure? And how would the firing rate (and neurons encoding said rate) be assigned to something novel like a blue dalmatian?

• Bradley Monakhos Says:
  Nov 09 2009 2:51 pm
  

  Thinking about - "So, how do you envision your system integrating new knowledge into an existing structure? And how would the firing rate (and neurons encoding said rate) be assigned to something novel like a blue dalmatian?" Formulating attack on the system you envision...

Leave a Comment