Pattern separation: What’s the problem?!
Author: Tim Bussey
HI Everyone. I wrote a short blurb on a blog site (http://timbussey.wordpress.com) with my (possibly hopelessly naïve) view about the problem with pattern separation – which is that there is no problem at all. Mike has encouraged me to post it on the site, so here it is (modified). Perhaps it’s a starting point for some discussion?!
——————————————————————————
What is pattern separation? I think this website provides as good a definition of it as any:
“The process of reducing interference among similar inputs by using non-overlapping representations.“
An often-used example of the kind of interference pattern separation reduces involves car parking. If I ask you about something you did 3 days ago, you can probably give me a good answer. But if you park your car in the same multi-story car park every day, and I ask you where you parked your car 3 days ago, it is exceedingly difficult. The memories of parking your car every day are so similar that they are difficult to discriminate, and become confused in memory. Pattern separation is process that helps to reduce this confusion – we’d be a lot more confused about all sorts of memories if we didn’t have it!
Seems straightforward enough. However in “the field” there seems to be considerable confusion about, amongst other things, how people “define” pattern separation, if indeed they do (see below), and how it is best studied experimentally.
I thought I’d write down my preliminary thoughts about this because actually — I don’t see any problem at all! From where I’m coming from, the study of pattern separation seems to me to be nothing new or out of the ordinary. So I am very surprised by the confusion. Let me try to explain.
I am a behavioural/cognitive neuroscientist; my degrees are in Psychology. In behavioural/cognitive neuroscience we have a basic paradigm for working. We postulate a putative process/construct/computation in the brain, e.g., working memory, attention, whatever. Then we devise tasks to try to capture that function, e.g., delayed response, target detection, whatever. We try to have parameters that we can manipulate, e.g., delay, duration of target. If, say, prefrontal cortex (PFC) damage leads to, say, a delay-dependent impairment in our delayed response task, we take this as evidence that the PFC is involved in working memory.
The behavioural pattern separation experiments — e.g., lesion the dentate gyrus, test on a putative test of pattern separation — are more of the same. As pattern separation putatively results in reducing the confusability, increasing the discriminability, of events, the parameter we manipulate is discriminability of events. There is nothing new under the sun here.
So when, for example, Adam Santoro writes that people, including me (Clelland et al., 2009, Science), define pattern separation as
“the literal behavioral ability to discriminate related stimuli”
and proceeds to argue against such a ‘definition’, I have no idea what he is talking about!
To return to the examples above, people who do those kinds of experiments on, e.g., working memory or attention are not defining delayed response as working memory, or target detection as attention. Those are just tasks, and they are using those tasks as assays of those putative processes/constructs/computations. (Of course one can always argue whether or not these are the right tasks to tap the constructs of interest, but that is a completely different issue.)
Now, having said that, Santoro is right in that some do seem to offer “behavioural” or “psychological” definitions of pattern separation — e.g. Hunsaker & Kesner — but I don’t really “get” that. There is no need for some separate behavioural definition of pattern separation. There are just tasks that we use to try to tap that putative function. Is this just semantics? I don’t think so; I think it’s important because talking about “behavioural definitions” will just fuel people’s misconception that there is something fundamentally different needed when studying pattern separation. But there isn’t — you don’t need a behavioural definition of pattern separation any more than there is a behavioural definition of working memory or attention.
So, What’s the problem?! There isn’t one.
Discuss … ?
27 thoughts on “Pattern separation: What’s the problem?!”
Tim.
No problem. Was nice to exchange in person at EBBS and I’m glad that one of the first public venues where some of the ideas on this board were discussed was that successful.
I’m curious though as to whether other folks on this discussion board can comment on this approach as well. Does it suffice to have a statement such as the one quoted by Tim above when discussing behavioral tasks and behavioral data? Do we believe that this would remove the confusion and immunize against misguided interpretations of our collective work? If not, what else should we be doing?
From my computational perspective, I think the statement is fine; but really I don’t know if it is necessary (or maybe should be necessary). I think everyone recognizes that there is a distinction between computational concepts and behavioral tasks.
What I would like to see is a recognition of something along the lines of the community recognizing that computational theory can show sufficiency, but almost never necessity (in a biological context), and behavior can show necessity, but almost never sufficiency. This works for both regional assignment of function (the dentate gyrus is needed to do pattern separation and is sufficient for pattern separation) as well as linking computation to behavior (pattern separation is necessary to do X behavior, as well as pattern separation is sufficient to do X behavior). There is a tendency of people on both sides to be a bit, shall we say, “sloppy” with their descriptions, and I think this leads to the frustration. And given that the behaviors and the models are NOT the same, one cannot simply combine the necessity from one and sufficiency from the other.
Put another way, if as a behavioral neuroscientist, conclusions were limited to “this region/process is required for this behavior to occur” as opposed to putting up a final figure with big headlines saying (young neurons -> Pattern separation; old neurons -> Pattern completion; we can all go home!) based on a single experiment, I think people would be happier. Not that anyone has ever done that particular example.
I wish no one has ever said that! I remember vividly when I was very guilty of imprecise language that was along these lines.
The manuscript was one wherein we lesioned dorsal and ventral CA3 and dorsal and ventral CA1 and compared the behavior of the animals on fear conditioning tasks.
We stated (I stated as I wrote the paper), that ventral CA1 learns… Whereas dorsal CA1 learns… And continued down that line for the behavioral results. I felt the reviewer was being pedantic at the time, but later realized they were right when they demanded I say instead that …animals with CA1 lesions were unable to perform…
It was an important lesson and I, to my knowledge, have always used more precise language in my manuscripts ever since. As far as I am concerned, this reviewer did me a great favor by teaching me this clear lesson.
I think Brad is right to ask that we use more precise language to describe what we are testing rather than using pattern separation or pattern completion as a catch all term to describe poor behavioral performance.
Thanks Mike. Have already added it to an article in proofs and one in revision!
Thanks
Tim
Tim,
This paragraph makes me oh so happy. I have started to use similar paragraphs in my papers as well and think it is excellent practice, so as to eliminate any potential confusion and not overstate the bounds of behavioral tests.
You have my vote.
Thanks Guys, It sounds like we may be approaching a solution whereby we can remain rigorous in our definitions, while at the same time keeping the work at different levels linked together by at least some key words. I am very keen to modify my behaviour to help this happen.
Using my own stuff as an example, I am thinking I could write my papers more or less as I have been, but when using the term pattern separation (*) provide a footnote or similar to make definitions clear. (And not appear ignorant!) How about this:
“(*) We are aware the term ‘pattern separation’ refers, in the original computational literature, to a specific proposed mechanism involving the transformation of an input representation to an output representation, in which the output is less correlated than the input [words stolen from Jim], resulting in non-overlapping stimulus representations (refs). Our behavioural tests assess the use of such representations. However it should be emphasised that our tests do not assess the mechanism of pattern separation, as defined by the computational modellers (refs), directly.”
What do you guys think?!
Thanks
Tim
To address Craig’s point about Jim’s definition, I think the problem is that it is almost fait accompli for many papers that pattern separation occurs in the DG, and while all of the data thus far are consistent with this account, the critical test is the one Jim proposes, an investigation of how much transformation of the input has actually occurred.. impossible to do without measuring the input. I think this is a problem because many editors and reviewers are so convinced that PS occurs in the DG (given the statements mentioned in papers across the board), they don’t seem to understand that this critical test has not been passed. So if someone proposes to do this as “the nail in the coffin” so to speak, there may be a “so what?” attitude that is due to the perception that this is either (a) already established or (b) not necessary. There is something terribly wrong with that.
Now onto the bigger issue at hand: I honestly don’t think there’s a problem in overuse of the term. I like using the term although I do tend to use it to describe the computational phenomenon and then everything else is “consistent with” said computation. I think it helps aggregate our literature under the same keywords which is helpful as has been previously mentioned. I wouldn’t want to censor anyone’s use of the term or tell people they can’t say the word, but it would be good to get our definitions and facts straight. It is a fact that PS in the DG has not yet been demonstrated. We have evidence for signals consistent with this account but absent the critical test, it still leaves something to be desired. This is important for folks to understand when they write about the process, so that its occurrence in the DG is not taken for granted, and thus the extension to neurogenesis for example would require a little bit more of working through the logic as opposed to the oft seen: pattern separation = DG, DG = neurogenesis, therefore pattern separation = neurogenesis.
I think the issues that we’re complaining about are not unique to our subfield in particular, but we have the strong advantage that we’re all talking about this now before the literature on this topic explodes. The subfield is small enough right now that we get to decide how to shape it. If we demand more careful definitions, so be it. If we demand certain caveats to be acknowledged, so be it. I just wouldn’t want this exercise to put a damper on anybody’s research efforts or make anyone think they shouldn’t use the term at all. All we’re saying is that the use of the term should almost come with a license of understanding the meaning and the caveats.
So my take on this is that we should just make sure that our papers always have operational definitions for whatever it is we’re talking about: hardly a novel requirement for scientists. We’re working on “emotional pattern separation” for example as one of our newest extensions and looking at amygdala-DG/CA3 interactions. You might wonder WTF is “emotional pattern separation”??? Well, I can tell you we have a very clear operational definition for how we’re defining these signals (you’ll have to read the papers when they come out to see if you agree or disagree). I could have called it “!$@#%$” for all I care, and as long as I define what I mean by that, everyone should be OK with it. I guess I’m less concerned about the misuse/overuse of the term and more concerned about appropriate operational definitions.
So as an author in this space, I promise to always have these definitions loud and clear in my papers. And I can tell you that as a reviewer of papers in this space, this is what I require. So if Craig wants to use “behavioral pattern separation” but he defines it in terms distinct from the neural or computational definition that would be fine with me.
My “it’s entirely way too late in the evening” two cents.
Glad to see all of this discussion and think Michael’s exhaustive review does a very good job structuring the pattern separation literature and formalizing the term.
Wanted to chime in from an outsider’s perspective in defense of Tim. I totally agree with Craig’s points in the other thread about the problems with ‘discrimination’. And I think ‘behavioral pattern separation’ makes sense intuitively to those that care enough to read pattern separation papers.
I can also tell you from my own outsider’s experience that use of the term ‘pattern separation’–while a bit unbridled–makes life easier on all of us. I did my PhD in an unrelated field and first heard of pattern separation 4 years ago. Partly the term itself inspired me to learn the concept and read a bunch more papers in the field (fun to see many of those authors in this thread!). Intuitively I came to understand ‘pattern separation’ as the cognitive ability to distinguish similar inputs, while ‘discrimination’ of dissimilar stimuli was handled by the brain differently. Yes it took dozens more papers for me to start questioning the blanket use of the term, but isn’t that true for any topic in any field?
So while I (and all of us serious scientist-types) agree with Adam’s general point to be careful about confusing the computational and behavioral definitions of pattern separation, I hardly think the potential confusion for the layman is hurting the field for those of us intent enough to be researching the topic. And as I’ve tried to indicate: I think the name has some useful cachet.
Jim wrote:
Nobody can disagree with this and Adam’s points come down to the fact that many papers aren’t talking about this. So, when talking about behavior, we’re never going to have this and I believe I’ve always acknowledged this point. One take is that when working at a behavioral level, the term “pattern separation” should be verboten as, well, we can never hit the definition above. Anyone who does… off with their heads! 🙂
Here’s my problem. Can any of us ever use the term then? Do the neural papers – things that are showing overlap metrics via IEG expression or via recording actually hit this definition? I’d say they don’t. So, they’re not allowed to use the term as well. Anyone who does.. off with their heads!
At which point, the only time the term can be mentioned, outside of the inky shadows of back room conversations is in computational papers. To me, that would be a shame.
To me, the rise of the term “pattern separation” has been a great thing. It’s gotten human researchers to think a bit more about representations and about computations. It’s gotten them to pay attention to computational models. When was the last time that happened? (Sorry, but… remember all, I got my Ph.D. with Jay McClelland doing models.) Instead of thinking that the purpose of the hippocampus is to do “recollection”, people are starting to think about what it is about the hippocampus that lets it play a critical role in recollection, episodic memory, or whatever. Sure, there are growing pains and things get munged, but I’d rather have that than the total disconnect we’ve had until recently.
So, I hope we can come up with a way of talking about this that doesn’t push “pattern separation” into the 100% accurate but rarely really applicable corner.
Craig
Tim: In response to your post, I feel like sharing a story here, since I think at some level we are all sympathetic to “pattern separation” as an interesting metric/function/behavior at one level or another. And I think most of us would like to see a way to more clearly map pattern separation across different scales, though I think we somewhat differ regarding what that would look like.
I was visiting a school (I won’t name it), and was talking to a neurogenesis researcher (not in this group), who as soon as I sat down picked up a paper (not by anyone in this group) in a very high profile journal that related neurogenesis and pattern separation and said “This is all your fault!” I spent the next half hour digging out of a hole. I don’t know how I got that reputation, considering that I have never claimed new neurons increase pattern separation and if anything my models don’t really support it. But I guess I may have been one of the first people to use the phrase in a neurogenesis paper (my 2006 time coding hypothesis paper), so it must be my fault? Never mind that the pattern separation – DG link has been around for over 30 years and so it makes sense to consider how new neurons may relate to that.
So there is something about pattern separation that annoys people. I don’t think it is simply the “neural correlate” request, because frankly there just aren’t that many systems or computational people who study population representations in this area. We’ve heard several other reasons here. My concern is that it is oversimplifies what is potentially a much richer function for the DG. That doesn’t make me hate pattern separation, but it makes me yearn for another dimension of analysis. That said, I think the trend towards designing behavioral tasks to look at the proposed pattern separation function is wonderful – far better than the previous status quo of doing water maze over and over again and figuring out a way to interpret it however one wanted. Ultimately, my goal as a modeler/theorist is for someone to test the resulting hypotheses at the animal level. That requires an ability to translate across scales.
Thank you Jim et al for this; these latest few posts have been clear and sensible.
Okay, you guys think the term “pattern separation” should only be used to describe a specific, putative mechanism at the neuronal level. It should not be used at the psychological/behavioural level. I hope I’ve got that right.
I guess this must mean, actually, that you don’t accept the definition provided on this web site, or definitions in terms of ‘conjunctive representations’, etc, because they are couched at the wrong level. Pattern separation may be FOR that, but that’s not what it IS. (But we’ve been discussing all of this under that definition on this website, so you can forgive a chap for being confused.) I hope I have understood correctly.
The trouble is that for me, I’ve been interested in conjunctive (non-overlapping) representations resolving interference (or ambiguity, whatever you like to call it) for about 2 decades. Indeed I think this concept is key to understanding cognition of all types, at all levels of stimulus representation. I’ve published a bunch of papers on this, and so have other people. We and others have the tasks that tap these putative conjunctive representations.
I’ve been working a lot on, e.g., perirhinal cortex, but of course we think the hierarchy of conjunctive representations we’ve been talking about extends on into the hippocampus, and we’ve said so in several places. My understanding was that people thought that happens (at least in the hippocampus) by a mechanism called pattern separation. So I talked about pattern separation in my papers. Seemed logical to me, and would help people who are interested in pattern separation to find my papers that are, I hope you agree, at least somewhat relevant to their interests.
But I could stop using ‘pattern separation’ and continue to use conjunctive representations and interference/ambiguity as I have for years. By doing that I will hopefully avoid reviewers demanding a ‘neural correlate’ – although I find that demand more than a little odd when you tell me that actually, no one has observed the neural correlate.
It is worth repeating that despite what some people have claimed, we’ve never defined pattern separation as behavioural discrimination. My first post included a discussion of this. The fact is that our behavioural results are consistent with a mechanism of PS, (even) as you define it. Consistency is not a very strong claim. However some on this forum have said I can’t even say that. I disagree, but at this point I just want to do experiments to find out how the brain works, and not get my papers rejected because I use the P-word.
I think what we are all interested in is how the brain works. We’re interested in function. I think our behavioural experiments have the potential to suggest the functional relevance of this putative pattern separation mechanism. However it is not clear everyone agrees, so maybe the right thing to do is play it safe and stop making that claim. I think that would be a shame though, because I think mentioning pattern separation in our behavioural papers helps connect up different levels of analysis. I agree that getting definitions precise is important. But I have this nagging feeling that in this case, we are in danger of throwing out a baby with the bath water.
Cheers
Tim
I want to chime in here and lend support to Adam and Brad. I think their arguments are spot on. As Mike knows (for we have had this conversation many times over coffee), I share the concern with the way that the term pattern separation is sometimes used in the literature that is not necessarily congruent with its original meaning from the computational literature. A few pennies’ worth of thoughts:
1) Formally, pattern separation is defined as the transformation of one input representation to an output representation, in which the output is less correlated than the input. One cannot directly test pattern separation as a function of a brain region without measuring the inputs to the brain region and its outputs.
2) This is extremely hard to do, and relies on assumptions of neural coding and representations that may be wrong.
3) I am not swayed that studying pattern separation is like studying working memory (or any other psychologically defined term). The scale of analysis argument holds here. Working memory is a psychological term that can be directly studied by behavioral tasks that attempt to tap into it. Pattern separation is (was?) a computational term about a neural network mechanism that might underlie the ability to store similar memories with reduced overlap. At this level of analysis, an analog to working memory might be the persistent activity of “delay” cells in the PFC (delay activity : working memory :: patterns separation: reducing memory errors). The computational/theoretical literature going back to Hebb and Marr envisioned such activity from reverberatory circuits, and the neural correlates were found by Fuster, Goldman-Rakic, and their labs (and probably others as well). It has been hypothesized that this persistent activity is a neural correlate of working memory. However, one would never (hopefully) design a working memory behavioral task and then claim that one is studying persistent neural activity. They are two different levels of analysis that may or may not be related to each other. Even if they are related, it is not true that all forms of working memory would necessarily be explained by this one neural mechanism. We don’t even know if persistent activity is a network activity as originally thought. Cellular mechanisms have been discovered, and they are in many brain areas. This type of activity thus may be a part of the toolbox of brain mechanisms used for a multitude of functions. The same is true of the relationship between pattern separation and memory interference.
4) This may be semantics or may be okay as long as we all know what we mean and insert the appropriate qualifying language in our papers. But I think it adds unnecessary confusion to the field, especially among those who are not as deeply embedded in these issues, including students. It is bad enough when we make the mistake of trying to tie neurobiological mechanisms directly to perceptual/psychological phenomena as if they were one-to-one mappings (LTP = memory; persistent activity = working memory). It is worse when we use the same terms to describe phenomena at different levels of analysis, as the terminology implies an identity that is questionable.
5) Ryan’s point that varying the input parametrically in a behavioral pattern separation task is a good one as it provides more confidence that such studies are likely to be measuring the behavioral consequence of a pattern separation mechanism.
6) Couching the arguments in terms of pattern completion rather than pattern separation may add another perspective on the same question. The theories that gave us the DG/pattern separation model postulated that CA3 does pattern completion. A number of studies have claimed evidence in favor of this. The classic task is to remove some cues and see if behavior and/or neural responses are affected. Normal animals are not affected by the cue deletion; manipulations that affect CA3 cause impairments. Therefore, goes the argument, this is evidence that CA3 does pattern completion. Here’s the problem: Without knowing what is going on in the inputs to CA3 (MEC, LEC, DG), one has no idea whether the deficit is due to an inability of CA3 to correct the errors in the corrupted/incomplete inputs from EC/DG, or whether the missing cues were “pattern completed” at any of the prior processing stages (retina, LGN, V1, V2, V4, TE, IT, perirhinal/parahippocampal, EC, DG) and the CA3 disruption caused some other processing deficit that impaired task performance. The same arguments hold if you substitute “separation” for “completion” and “DG” for “CA3”.
7) We just submitted a manuscript with DG and CA3 recordings that, in combination with our previously published entorhinal recordings, we argue provides direct evidence for DG showing pattern separation and CA3 showing pattern completion/error correction (DG outputs less correlated than its inputs; CA3 outputs more correlated than its inputs). Will be interested in how the results and arguments are received.
I agree with Jim entirely. Pattern separation is a absolutely a circuit level phenomena that plays out at the level of cell populations rather than at any single cell (so far as we know at least).
The difficulty situation we have ourselves in is that we are trying to answer what seems like a simple question: “what is pattern separation”. So far as has been proven, at this point in time pattern separation is little more than a mathematical and computational construct that comes out of the end of a computational or theoretical model. We have a lot of candidate mechanisms whereby this process may occur, but we can in now way be sure if we are correct until we can actually acquire data from all the cells along the EC-DG-CA3 circuit so we know for certain that similar patterns input to the DG from EC are in fact decorrelated at the level of the mossy fibers as the models suggest. The simple truth is that we have no proof for this yet at a neurobiological let alone psychological level. As stated above, Jim may actually give us a first glimpse at a cellular population correlate to pattern separation (I am already cheering and hoping it gets fast tracked for publication so I can read it).
On the other hand, just to be troublesome, laser focusing our attention on the dentate gyrus may be causing us to miss the forest for the trees. It has been demonstrated that nonDG structures, even within the hippocampus, can show behaviors that are rather similar to the “pattern separation” we talk about in the dentate gyrus (referring to CA1 of course and “temporal pattern separation”).
Similarly, it has been clearly demonstrated that nonhippocampus structures, (e.g., the amygdala) show similar types of pattern separation-y effects, but at a courser resolution than the spatial and visual object pattern separation we are all used to thinking about.
I’ve been thinking about this conversation a bit more the last few days; and I believe that some of the friendly acrimony (is that possible?) across some of the responses is 1) an indication that there a problem, but 2) no one really knows what it is – at least not well enough to communicate it. And as Tim says, we are all frustrated by having to satisfy multiple reviewers who have conflicting interpretations of the words that we’re using.
Let me throw a possible explanation. Computationally, I believe that the DG has multiple roles in episodic memory formation. The primary function is conjunctive encoding – taking representations from different cortex modalities and unifying them into a single encoding pattern for CA3. There is substantial evidence for this – anatomical, physiological, behavioral, and computational. There likely are secondary functions, say something like neurogenesis-induced temporal coding or something like that (perhaps a “when” with the classic “what” and “where”). Further, the DG is responsible for “burning” new attractors into the CA3 network. To be effective, those have to be separated from one another. In this view, pattern separation is not a function, per se, but rather a design constraint. Whatever the DG does, it has to preserve a level of orthogonality of its outputs (or more precisely, downstream CA3 memories).
On the other hand, increasingly behavioral studies in rodents and humans are looking at how information is encoded (separation) and less often looking at what information is encoded (conjunctive encoding). This is still useful, because the DG is doing pattern separation. But is it the function or the constraint of the system that is being tested? Maybe this is simply semantics; but as I’ve argued in the past, the best way to separate two events is to not encode one and encode the other – a 0 or a 1 digital separation. So clearly separation alone isn’t sufficient.
Adam,
I think the issue of input similarity is a good one but you may be too harsh on folks here. I think everyone at least on this forum fully understands that it’s the neural inputs that are relevant and that the goal of the stimulus input is to attempt to get some correlate in neural inputs. That’s the logic of 99% of our experiments. This is not a contentious point at all, at least among those of us who do this kind of work (although I cannot speak for everyone).
I think the larger point which we make in our TINS review and you do so in your Frontiers paper as well is that in order to infer that a pattern separation computation occurred, consideration of the input and particularly the degree of overlap in the input patterns is critical. That is the ultimate test of whether the process occurred.
Hi All, I feel I should reply to this, but I only have a sec, so with profuse apologies for brevity! :
>one can interpret “similar inputs” as referring to “similar stimuli.”
1. I don’t think you have to worry about this. People realise that stimuli are outside of the brain.
2. Yes I think we do assume that similar stimuli activate some of the same neurons. I would have thought that was a reasonable assumption.
I think I am going to take a break from this for a few days now. Enjoy!!
Tim
>pattern separation IS the neurobiological correlate (of discrimination).
So maybe there is a problem after all, which for me crystallises in this statement. I think there may be two cultures here, one that thinks a statement like this is acceptable, and one — perhaps those of us schooled in psychology/behaviour? — who think the neural phenomena we observe may cause pattern separation, i.e., “The process of reducing interference among similar inputs by using non-overlapping representations“, but that that remains a _hypothesis_.
Which is all fine in a friendly discussion like this; there are plenty of similar cultural/paradigm divides in neuroscience. But practically, it means that we do 10 experiments manipulating molecular mechanisms in the DG, using (what I think are precise and well-controlled) behavioural tasks, and it gets rejected from every journal because we don’t have a ‘neural correlate’ which according to the reviewers IS pattern separation. I don’t think that’s right. And that is the closest I am willing to come to a sour-grapes-sounding rant!!
But here’s something I agree with, from Adam
>in fact pattern separation is a ubiquitous computational phenomenon occurring everywhere in the brain
Yes! I think I’ve been talking about object-ish-level pattern separation in perirhinal cortex for a decade and a half — just not in those terms (although Kesner, Hunsaker etc have). DG is involved in spatial pattern separation. (Although I also don’t think words like space and object map neatly onto the brain either; why would they.) That’s probably another discussion, and I suspect other contributors on this site would have something to say about it! 🙂
Thanks for listening … again
Tim
Tim,
These types of statements really confuse me: “…who think the neural phenomena we observe may cause pattern separation, i.e., “The process of reducing interference among similar inputs by using non-overlapping representations“, but that that remains a _hypothesis_.”
There are a lot of terms in here that are ambiguous, and have precise meanings depending on the reader. Here is my interpretation:
“Similar inputs” should refer to the degree of overlap in cell population activity upstream of region X. It should not refer, and never has referred (in the computational work) to similarity of stimuli presented to an organism. Thus, we cannot speak of pattern separation as the use of non-overlapping cellular representations for similar stimuli, since this is not how pattern separation works as a mechanism.
One can assume that similar stimuli presented to an organism in some task result in similar cell population activities upstream of the pattern separating region, but this is just an assumption. This has never been shown, and it is not at all trivial. In fact, this demonstration is crucial to the validation of the pattern separation tasks we use.
On the other hand, one can interpret “similar inputs” as referring to “similar stimuli.” If this is the case, then pattern separation is being defined completely differently. This is the crux of the issue. One must demonstrate a neural correlate when speaking of pattern separation in this sense – that is, one must demonstrate that these similar inputs produce similar population activity, which is then parsed into dissimilar population activity by the pattern separating region. Again, this has never been shown.
I think Brad’s comment hits the nail right smack on the head. I particularly like the following points:
“To me, saying that “process X contributes to pattern separation” is actually quite meaningless for my purposes at a cellular or population level, because to some extent every neuron, whether individually or as a population, separates some things and generalizes other things. …Maybe the DG does it a bit more than everyone else, but that doesn’t mean it is the DG’s reason for existence.”
I think this is a crucial point. It is an assumption that the tasks we use to assess pattern separation are somehow dependent on cellular pattern separation in the DG, when in fact pattern separation is a ubiquitous computational phenomenon occurring everywhere in the brain.
“…but I am comfortable saying that the tasks being described simply do not help us decipher what the underlying computational mechanism is.”
I agree with this wholeheartedly. Hence my reluctance to even state that these behavioral tasks are “consistent” with a pattern separation mechanism. They’re also consistent with LTP, but we don’t feel the need to mention that. In my opinion, pattern separation needs to be thought of as a member of a computational toolbox used by the brain to process information. The kinds of statements Brad yearns for seem esoteric to us studying behavior, but they are precisely the correct ones that need to be made.
“I think the ultimate success here would be to recognize as a community that what passes for separation at a behavioral level may simply be a dependence on DG involvement in encoding – irrespective of population representation separation – at a computational level. ”
Precisely. Hence, the conclusions to be made from these tasks (whether you want to call them “discrimination tasks”, or any other name) is that the the task involves some form of encoding in the DG. No more. The conclusions should not be that these tasks involve, or even are consistent with pattern separation. I take a strong stance in saying that we should not mention this “consistency” because I believe this wording just causes confusion for those that are not deeply entrenched in the topic. I’m not married to this opinion though 🙂
Tim: you ask “Is it really appropriate to demand a “neurobiological correlate” or else you’re not studying pattern separation?”
I think on this one I’d have to answer a definite yes. The reason: pattern separation IS the neurobiological correlate (of discrimination). So when studying discrimination, to claim that pattern separation occurs is to claim the existence of a neurobiological correlate of discrimination. Existence cannot be claimed if it is not measured. We don’t do a working memory task and claim that LTP increased, for example, without measuring LTP.
“But, to use the WM analogy, if you have a PFC lesion and a validated working memory task, but no lower-level (e.g., neuronal) correlate, it doesn’t mean you are not allowed to say you are studying working memory”
The equivalent analogy does not question whether you are allowed to study working memory. Of course this is fine. The equivalent analogy implies that you are not allowed to say you are studying protein synthesis, or LTP, or mRNA splicing when you are only doing a watermaze task. Watermaze is the task that captures working memory. The purpose of the watermaze is not to capture protein synthesis/LTP/mRNA splicing. Similarly discrimination tasks should not be thought of as a capture of pattern separation.
Ha! I’m sure we’ll find something else to bicker about!
Working Memory — the whole time I am thinking “maybe no the best example” because of exactly what you say!
Cheers
Tim
Tim — agreed — so we should all just agree to not make specious reverse inferences. But is this really the only argument here? If so, it’s not much of an argument and we really need to find something else to bicker about 🙂
Now, that said – as the link gets stronger and stronger between particular behavioral performance and signal in / operation of a region (say, hypothetically, performance in our BPS task and the DG) there certainly is the temptation to use the behavioral performance as an “assay of function”. This is, of course, the cornerstone of neurology.
FWIW, I’ve called this task many things. It’s still called in many directories on the server the “Mnemonic similarity” task. For awhile it was called the “SPST” or “Stark Pattern Separation Task”. That could have been read egotistically and I’d meant anything but that — the notion being there would be many out there from others and this might disambiguate it. That did imply I was a one-trick pony too and so that version got nixed.
But really — so long as we define what we’re talking about and be good scientists and not succumb to invalid inferences, what’s the big issue? On that front, I’m clearly with you Tim…
(As for “behavioral working memory” – the term “working memory” has gotten pretty bloated itself and means totally different things IMHO in the human and animal literature.)
>Just because a dentate knockdown for example causes a deficit in discrimination behavior in a “pattern separation task” doesn’t mean that whenever you have a deficit in such a task it is because there is a DG problem
Agreed — just as you can’t say a working memory deficit necessarily means PFC dysfunction. But in practice does this really happen – does anyone really say “My patients are impaired on pattern separation and therefore they have a dysfunctional DG” without appropriate qualification?
Another question: Is it really appropriate to demand a “neurobiological correlate” or else you’re not studying pattern separation? Sure, it is always nice to have, e.g., the behaviour of neurons as well as the behaviour of the whole animal, no matter what process you are studying. But, to use the WM analogy, if you have a PFC lesion and a validated working memory task, but no lower-level (e.g., neuronal) correlate, it doesn’t mean you are not allowed to say you are studying working memory…
Tim, regarding your question as to whether or not the reverse inference issue happens. The answer is YES!! Many of these papers end up in my hands and they end up rejected but some of them get through the filter and do get published. A recent example I have seen (without mentioning names) went as far as to say that impairment on a PS task in disease X was evidence for DG impairment and specifically with neurogenesis, especially given that there is some evidence in animal models of disease X also show DG neurogenesis loss. It’s more rampant than one might think and especially in the rapidly growing clinical literature.
Regarding your second point, if you’re using a task that you’ve previously validated I see no problem making some link to neurobiology across papers and not within the same one. However, we see that there are many caveats to task design. Heck, even something as simple as incidental instructions, vs. explicit instructions can make a difference, timing, number of stimuli, etc… All of these things can change a task in such a way that deficits become inexplicable in the absence of additional neurobiological validation.
Since we’re opening up and revealing our deepest, darkest thoughts on pattern separation, I thought I’d give my perspective from the computational neuroscience side of things.
We’ve had two proposals for what the problem is: 1) there isn’t a problem, and 2) there are potential logical fallacies in its application.
I’m going to give a third, which hopefully illustrates why I can agree with much of the last two proposals and also see there being something to work on.
The problem: “Pattern separation” is currently used, and defined, in different ways by different people with fundamentally different questions. Simply put, the problem is scale of perspective.
The solution: We need to pattern separate “pattern separation” as a community.
First, why is this even a problem? Tim brings up the point that this is the case with working memory and attention. I think these are hardly pillars to hold up as positive examples; as the ability of these terms to mean whatever someone wants them to mean is at best frustrating and at worse detrimental. This inability to define cognitive processes independently of scale has, in my opinion, contributed our ability to integrate cognitive brain regions (the more interesting ones, in my opinion) into the fairly rigorous approaches developed within systems neuroscience. Given that those terms are ingrained into the community, they aren’t changing. But we’re still early enough in the game to affect whether “pattern separation” becomes “working memory” or becomes something better.
I’ll give a try, but I too have my own biased perspective:
As a computational neuroscientist, I care about how a system, say the hippocampus, processes information. It isn’t sufficient to put a big box around it and say “it forms episodic memories, stores them short term, and ships them off to cortex for long term purposes”. I want to know how it does that; not just for kicks, but to be able to gain intuitions of either therapeutic interventions at the cellular level or new algorithms for Google to use or something.
To me, saying that “process X contributes to pattern separation” is actually quite meaningless for my purposes at a cellular or population level, because to some extent every neuron, whether individually or as a population, separates some things and generalizes other things. Notably, this is one of the main points in O’Reilly and McClelland’s seminal modeling paper on pattern separation back in 1992; separation can be seen in every hippocampal region. Maybe the DG does it a bit more than everyone else, but that doesn’t mean it is the DG’s reason for existence.
Rather, I yearn for statements such as “The dentate gyrus creates nearly orthogonal attractors on the CA3 recurrent network, thus minimizing interference between memories” (rephrase of Treves and Rolls, 1992) or “Neurogenesis contributes to an increased resolution, or level of detail, of memories by altering the DG’s representation of cortical information” (rephrase of my Aimone et al 2011 paper) or “Place cells in the DG will rate remap earlier than downstream CA3 neurons” (rephrase of Leutgeb 2007).
Now, I can appreciate that these mechanistic statements are not of much value to behavioral and cognitive neuroscientists, as examining decorrelation of population representations is effectively impossible using fMRI, and cognitive or animal behavioral tasks, and is a stretch even in vivo animal physiology techniques. What can be measured are tasks of discrimination and looking at overall regional involvement. And as I’m not at that level, I am loathe to tell higher level scales how to define their terms, but I am comfortable saying that the tasks being described simply do not help us decipher what the underlying computational mechanism is.
My goal from the computational perspective is to help give you better tasks that could potentially discriminate (ha!) between the different potential computational mechanisms at play. For instance, at the computational level, we have quite the debate about what young neurons actually do. The answer here is not “they increase pattern separation”, it is something along the lines of “preferentially drive interneuron activity” or “provide broadly tuned representations to CA3” or “sit their quietly learning what to respond to later as good little school kids should”. All of those mechanisms could impact separation.
I think the ultimate success here would be to recognize as a community that what passes for separation at a behavioral level may simply be a dependence on DG involvement in encoding – irrespective of population representation separation – at a computational level. Likewise, from the computational perspective, it is necessary to embrace that computational mechanistic effects are merely trivialities unless they can map up into some behavior. Frankly, I’m not going to be satisfied with any DG definition until I get a good theoretical model for how it affects the CA3 and CA1; only then maybe could I see how it ties into the behaviors tested so well by others.
Lastly, let me revisit what I started with and say that compared to older communities, we in the DG/PS field have a unique opportunity to bridge the molecular to computational to systems to behavioral to cognitive domain. The split in hippocampal researchers between spatial-processing obsessed place cell physiologists and declarative memory based cognitive neuroscientists is unfortunate. Similar for many other domains. Let’s get this right.
This is great, it succeeded in distracting me from monday morning’s urgent laundry pile of things to do:)
I agree with Tim and also on the need to, in as much as is possible, support behavioral analysis with circuit correlates (remapping, lack thereof as seen in CA3 rigity-Tanila et al (rats), Yassa et al..(Humans) that formed basis of early theory. Contextual disc. learning which I presume Yassa is referring to, doesnt capture tight modulation of paramatric features (other than varying similarity across 3 different contexts), but unlike, other tasks (touch screen, object disp), is supported by circuit correlates such as global remapping (Niibori et al….Deng et all…). Ofcourse, you cant use a “pattern separation task” and conclude that the DG-CA3 circuit is impaired without providing causal linkage with circuit manipulation. Now, back to laundry….
Thanks for starting this thread Tim and Yassa, for this blog. It is incredible how much attention this phrase has received, and with attention comes scrutiny and introspection, which is good for science.
I couldn’t agree more with both Ryan and Tim but let me perhaps try to reframe this and see if this is the issue that is at the epicenter of the debate. I have no problem at all calling something a pattern separation task as long as it is clear that the task is used in the same way something like “hippocampal task” is used. Tasks are used as assays of certain processes or regional functions. That’s all fine and good. In my opinion however we have a severe reverse inference problem that manifests quickly. Just because a dentate knockdown for example causes a deficit in discrimination behavior in a “pattern separation task” doesn’t mean that whenever you have a deficit in such a task it is because there is a DG problem, especially if now we are talking about a new disease or population. Worse yet is using discrimination deficits to say something about neurogenesis for example without testing anything neurobiologically. I think the latter is circumstantial evidence and in this case the tasks should not be called pattern separation tasks. When we observe for example rigidity in ca3 and see that it maps onto a behavioral discrimination deficit we have a little bit more leverage to call it a PS task because heck, that’s the neurobiological correlate. Absent such correlate, I’m not sure we can do the same. This also applies to pattern separation behavior, etc…
Thoughts?
I think you raise a great point here. As one of the people who has been drawing ever smaller boxes around definitions (ie Kesner and Hunsaker) I can see there is a limit to this and we may be encroaching upon it with pattern separation.
Perhaps a better way to describe the behavior is to look at behaviors that reflect underlying pattern separation processes. That way saves me from having to get into very long twitter discussions trying to define when discrimination does and does not reflect pattern separation.
I know for one, I am leaning this way since we are getting perilously close in my opinion to making the throat tern separation mean something very different than what we are styling by forcing it into increasingly Procrustean boxes.
Even as recently as a year ago when writing our latest review I had much stronger views others definitions…I think I had just not looked at the behavioral data without re-interpreting their graphs through the filter of my computational/theoretical collaborators. Upon reflection, I think you are right in that so long as we design tasks that parametrically modulate interference/discriminability we are saying the exact same thing!
Again, great post!
Comments are closed.