Talk:Models of synaptic plasticity
"Theoretical analysis indicates that not only Hebbian like synaptic potentiation is necessary but also depression between two neurons that are not suffciently coactive." This should be expanded: necessary for what? (to prevent network instability I assume?)
"1) Phenomenological models: These are very simple model" (sp. models)
"Phenomenological models" I think reference should be made to the Dayan-Abbott book where there is a more comprehensive overview of phenomenological models. " such as in PCA or BCM models." I do not personally understand the link to Principal Components Analysis (PCA) here and think this should be explained or replaced with another example. Although the PCA section is not yet written, I highly doubt they will refer to PCA models of synaptic plasticity to explain this point.
"Spike timing based models":
Maybe refer to as an actual example of a more general "temporally asymmetric Hebbian" model that depends on order of spiking and incorporates causality rather than simple rate correlation (eg Blum and Abbott 96).
"Biophysical models of synaptic plasticity":
" A simples choice (Karmarkar and Buonomano, 2002)" (sp. simple) "pairing indice plasticity" (sp. induced) "transients induced by paring" (sp. pairing) "and therefore signifficantly effect" (sp. significantly) "is unable to account for why LTD s induced" (sp. is)
"To account for such LTD a BPAP with a wide tail potential is either explicitly assumed, or implicitly included by the parameter choice of the postsynaptic neuron." I didn't know this was true. Please give an example by citation.
"Calcium based models that account for STDP typically result in another form of LTD at \Delta t > 0, which in a consequence of the continuity of the magnitude of Ca transients as a function of \Delta t (Fig. 4)." (sp. in -> is)
I think this should be expanded, noting the desire for a single Ca-dependent mechanism to account for pairing and STDP, referring to Fig.2a. Then add eg: "That is, if Ca influx is large with small +ve \Delta t and zero with very large +ve \Deltat t, then for moderately large =ve \Delta t Ca influx must pass through the range of moderate increase that would induce LTD." Then reference the suggestion of a "veto mechanism" by Rubin et al. as well as your work on stochasticity. Personally I think the veto mechanism is a lot more relevant (so should be discussed) than the mechanism suggested in Fig.2c (and discussed in the following paragraph) which is not supported by experimental data as far as I know.
" has not yet been rigouresly modeld" (sp. rigorously modeled)
Figure 4: "a. Some experimental results indicate " I think "most experimental results indicate ..."
"b. Other experiments find " Some experiments, in HC slices but not cortex, find ...
"single experiments (+) diverge significantly from" Rather "have significant variance around ..."
Modeling the signal transduction pathways associated with synaptic plasticity:
"Under certain conditions, such enzymatic models enzymatic models can be" enzymatic models (once only!)
"In addition, such models assume we know the key molecules and kinetic coefficients of their interactions, assumptions that might not currently hold (Castellani et al., 2005)."
But your example, Eq.(4) assumes no such thing! This would be an even greater problem if stochastic modeling between compartments is carried out, since, for example, if average concentrations of molecules are unknown, their particular spatial distributions and fluctuations provide even more unknowns.
(I assume my review goes here)
Overall the article is well structured and understandable and I have only a couple of suggestions:
a) Like reviewer 1, I am also confused about 'PCA models' ... this cannot refer to principle component analysis, but I am not aware of any model of synaptic plasticity with this acronym.
b) An original citation for the BCM rule would be appropriate
c) In 'Spike timing based models', first paragraph: 'Under different assumptions about ...' -> 'Under certain assumptions about ...' ?
d) I thought the discussion was a little short on the phenomenological models, there has been quite some work on additive versus multiplicative rules and the general question of 'run-away' potentiation that may be worth mentioning
e) I took the freedom of fixing up clear typos in the text myself, many of them the ones mentioned by reviewer 1, but also a few in the references; I noticed that the Linsker 1989 citation was ambiguous, please check that I edited what you meant.
f) Figure 1i shows a discontinuous STDP rule which is fairly wide-used. Personally I would suggest to also show one of the continuous alternatives (which are preferable from a theoretical point of view because of stability issues)
g) Last sentence before 'Modeling the signal transduction pathways ...': 'Although such a model can account for STDP, it is unclear how it can account for pairing induced plasticity': I do not understand this statement. What is STDP if not pairing induced?
One more review:
I think that the article is correct, clear and well written and I have a few specific remarks (see below). In general I believe it is not easy to write a review about models of synaptic plasticity given that the literature on the subject is vast. The author focused mostly on a certain class of biophysical models and he certainly did a good job in reviewing them. The parts that should be complemented with a few references are those related to spike-based simplified models of synaptic plasticity. There are several works in which the authors analyze the computational role of synaptic plasticity in different contexts. For example:
1) STDP and temporal coding (Gerstner, Kempter, van Hemmen, & Wagner, 1996) 2) how a modified version of STDP can improve the performance of a visual perceptual task (Adini, Sagi, & Tsodyks, 2002) 3) STDP for the emergence of direction-selective simple cells (Buchs & Senn, 2002; Senn & Buchs, 2003), Song et al., 2000 is already cited 4) STDP is used to encode simple temporal sequences in Rao and Sejnowski (2001) 5) In (Gutig & Sompolinsky, 2006) the principles of the perceptron learning rule are applied to the classification of temporal patterns of spikes 6) Fusi, Annunziato, Badoni, Salamon, & Amit (2000) and Mongillo, Curti, Romani, & Amit (2005): the authors study a spike-drive synaptic plasticity model to learn attractors 7) Abbott & Nelson, 2000: this is one of a series of works on the regulatory properties of STDP
They might belong to the article on STDP, but then I'm not sure what the article on synaptic plasticity should be on. In such a case the author should probably extend the discussion on the consolidation mechanisms underlying the expression of long term modifications and change the title into "Biophysical models of synaptic plasticity and memory consolidation". These remarks are more related to the general organization of the articles than to the specific contribution of Shouval.
Specific remarks:
a) Rate based models: there are several of these models in the literature. I'm not sure about who was the first to introduce a covariance based learning rule, but I believe that Terry Sejnowski should be certainly cited (he has a famous paper published in the seventies). Learning prescriptions like the one of Rosenblatt or Hopfield are also based on the same principle. Again, I'm not sure whether the last two belong to this article, but Sejnowski's paper should certainly be cited. I also agree with the second reviewer that it is important to add a reference to the original BCM paper.
b) Spike-Timing based models: the author should probably discuss the limitations of the models based on spike-timing (he partially discusses a much richer phenomenology in what follows). Most of the LTP/LTD induction protocols cannot be reduced to a synaptic dynamics which is based only on spike-timing. For example, Jesper Sjostrom and colleagues (cited in the article) showed that there is clearly a dependence on the subthreshold depolarization of the post-synaptic neuron. Moreover STDP is observed only in a limited range of firing frequencies of the pre and post synaptic neurons (again, Sjostrom et al). I believe that any biophysical model should capture the whole, rich phenomenology observed in the experiments.
c) Biophysical models of synaptic plasticity: the author describes mostly his own results. I think it would be fair to specify at the beginning of this section what kind of long term modifications and what experiments he is going to reproduce with the proposed model. The literature on LTP/LTD is vast, and so is the literature about the models. The author has in mind specific protocols, of experiments performed in specific areas of specific animals. It would be fair to warn the reader about the richness of the phenomenology observed in experimental works. It is also important to clearly discuss the differences between presynaptic and postsynaptic long term modifications.
d) The author speaks several times about Back propagating action potentials. It would be fair to mention that there is an ongoing debate on the subject (see for example Lisman Spruston, Postsynaptic depolarization requirements for LTP and LTD: a critique of spike timing-dependent plasticity, Nature Neuroscience)
______________________________________________________________
I am not sure where I respond to ther reviwers, I guess here. _______________________________________________________________
I have made signifcant changes to the article based on these reviewes. Below is a response to the reviewers and an explanation of what changes I did and did not make.
General response
First I would like to thank the reviewers for a constructive set of comments.
There was a general concern, primarily by the third reviewer, regarding, what this article is about? Indeed, the title is very broad and can be interpreted in many ways. I would have preferred to call this: "Biophysical models of synaptic plasticity". However, Eugene preferred "Models of synaptic plasticity". I decided to stress the biophysical models because other articles in scholorpedia cover STDP, BCM, PCA (yes, PCA).
This article is supposed to be short (~2000 words), and with a small number of references (~25). Many of the comments wanted me to add text or references. I often did this, but as a results the article is longer and has more references. Many of the things you asked to include are things that were in there originally and taken out to reduce space/references. However, since I still want to keep it short, and since there are other somewhat overlapping articles, I decided not to elaborate much more on the phenomenological models. I think the suggestions to add references regarding STDP papers are good, but this should be in the STDP article.
Reviewers 1 Expanded a little on why depression is necessary.
Dayan and Abbott book reference added.
PCA - By PCA I refer to the Hebbian model with decay proposed by E. Oja in 1982. This is commonly refered to as the PCA model. However, two reviewers did not know what I mean, so I clarified my intention. By the way, it seems that Oja is the author of the PCA article of scholarpedia, so I think he will discuss his neuronal model.
"Maybe refer to ... I am not sure what this comment is about, I have referred to several models with temporal asymmetry. Some of the papers referenced try to reduce the asymmetric STDP rule to a rate based co relational rule, but they start from a spike based rule. In any case I think that there is a whole article on STDP, and it will have room to discuss this in much more detail.
"To account for LTD ..." I didn't know this was true ...
I am not sure what the reviewer means here, did not know what was true, that there is a wide tail to the BPAP, or that some models implicitly assume a wide tail? If it is regarding the first I can site papers that show a wide BPAP, although this might not hold in all cells. Regarding the second I can add references of papers, for example Abarbanel et. al. Kitajima and Hara which used an HH type model to generate a spike.
"I think it should be expanded .... I expanded.
"Figure 4. Some experimental results indicate" I think most experimental results. Here I beg to differ. Papers in Hippocampal slices, the most used preparation for synaptic plasticity, clearly show an LTD region for \Delta t>0. I several additinal papers in other systems including Bi and Poo 98 (Culture), and in the Froemke and Dan papers (VC) and Sjostrom papers there were data points at positive \Delta t which produced LTD. This data was sometimes fit with an exponential, which obviously cannot produce LTD. Yet other experiments did not put a significant number of data points, or any data points (Markram) in this region. These LTD points might indeed be due to some noise in measurement, however this is not yet clear. Therefore, I think some is more correct and imposes on the data less preconceptions.
Similarly I think "single experiments diverge significantly" sums the experimental results correctly.
"In addition such models assume ..." I agree with this general comment, and I changed this paragraph to reflect these comments.
Reviewer 2
a)PCA - see response to reviewer 1
b) I added an original citation for BCM
c) Accepted, thanks
d) As described above, I did keep the phenomenological models discussion short, because of space limitations and because other articles cover these types of models (BCM, PCA, and part of STDP).
e) Thanks
f) I can add a continuous curve to 1i, but I think this will be better paced in an STDP article.
g) In the experimental literature, as well as in Fig 1d-f of this article a pairing protocol is one in which the postsynaptic cell is depolarized while a low frequency stimulus is presented presynaptically. However, since this terminology is indeed confusing, I changed the text in this sentence.
Reviewer 3
Regarding the general comments see above.
specific comments: a) I added a reference to the original BCM (1982) and to Sejnowski (1977). Again, I had these papers in the first drafts, but cut them out to reduce the number of references. I hope the editors do not know how to count :)
b) Limitation of spike timing based models. I added a paragraph to reflect these concerns at the end of the spike timing dependent models subsection.
c) There are several concerns here. The first is that I talk mostly about my own work. I think I cite other similar work as well, and bring my work as an example of this approach, but I do cite other similar works. In the new version I cite more. Second "... specify what kind of long term modification ..." I agree that there are many types of long term plasticity, that differ between systems, layers, specific synapses. In a research article it would eb best to account for one type. However, since this is a broad review, I do not want to restrict myself to one system etc. I included new text in the paper to explain these points. Third, pre vs. post site of expression. This is very important, but I don't think I have space to do this issue justice and in addition this is a review of models, and models of plasticity have to the best of my knowledge not addressed this issue explicitly.
d) There is plasticity that does not depend on back spikes, for example pairing induced plasticity, and possibly low frequency induced LTD. In addition there might be local spike events that do not originate in the soma. Biophysical models have indeed modeled some plasticity in the absence of back spikes, and it will not be difficult to model plasticity in a compartmental model due to local spikes. However, 1. I do not have much space for this. 2. I am not sure if there are published models of plasticity due to local spikes, and remember this is primarily a review of models. However, I did include one sentence about this at the end of the spike timing dependent plasticity section, and another note when in the calcium dependent models section.
Reviewer 1 again: I'm happy with these changes -- nice article -- I'll hit accept but request you please fix the typos using a spell-check and make sure each sentence starts with a Capital etc. :-)
Reviewer 2 again: Shaped up nicely ... and I actually learned something new about PCA.
