By: Adam Santoro
The following text is my personal view on the pattern separation problem. In my opinion, the problem stems from the use of logical fallacies. I don’t accuse anyone of using any of the following fallacies, and, to be honest, most of it comes out in verbal discussions and not in text. I’m also ignoring any differences in definitions of pattern separation, as I think those are not part of the “main problem.” There are a lot of subtleties in the text below, but I think they are crucial. So…here’s my take.
I’ll start by pointing out an important logical fallacy called “affirming the consequent.” It has this form (taken from Wikipedia):
As modus ponens, the following argument contains no formal fallacies: 1.If P then Q 2.P 3.Therefore Q A logical fallacy associated with this format of argument is referred to as affirming the consequent, which would look like this: 1. If P then Q 2. Q 3. therefore P This is a fallacy because it does not take into account other possibilities. To illustrate this more clearly, substitute the letters with premises. 1. If it rains, the street will be wet 2. The street is wet. 3. Therefore it rained. Although it is possible that this conclusion is true, it does not necessarily mean it MUST be true. The street could be wet for a variety of other reasons that this argument does not take into account.
Let’s use a more science-y example:
1. If neurons are exposed to high levels of mystery growth factor MYST-A, then dendritic arborization increases 2. There is increased dendritic arborization 3. Therefore the neurons were exposed to MYST-A
This logic, of course, would never pass reviews. However, the following is logically valid:
1. If neurons are exposed to high levels of mystery growth factor MYST-A, then dendritic arborization increases 2. Neurons are exposed to high levels of MYST-A 3. Therefore there is increased dendritic arborization.
Whether this is true is a different question from whether it is valid. The above conclusion is only true if premise 1 (and 2) is true. We’ll get back to this point soon. Let’s go back to the fallacy.
Pattern separation is a phenomenon at the level of the cell population. It is a cellular mechanism. This is the logical form that affirms the consequent (i.e., it is invalid):
1. If pattern separation occurs, then the animal discriminates. 2. The animal discriminates. 3. Therefore, pattern separation occurs.
This is much different from this valid logical form:
1. If an animal discriminates, then pattern separation occurs. 2. The animal discriminates. 3. Pattern separation occurs.
The first form affirms the consequent. It does not take into consideration any other possibilities to explain successful discrimination. For example, the animal could have discriminated because of some other mechanism, let’s call it the “discrimination mechanism” (DM). By illustrating premise 2, we do not know whether PS or DM is at play. This is similar to seeing an increase in dendritic arborization from the first MYST-A logical form – do we know that dendritic arborization was caused by MYST-A? How about some other growth factor?
Now for the second form. The second form is logically valid, but the truth of the conclusion depends on the truth of the premises. For our purposes, premise 1 is crucial:
1. If an animal discriminates, then pattern separation occurs.
Here is the excellent example Tim Bussey uses in his blog post:
“The classic 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!”
Bussey explains that PS can reduce interference, essentially increasing discrimination. For us to confidently (and truthfully) state that “if an animal discriminates, then pattern separation occurs” we need to know that discrimination necessarily entails pattern separation. By “necessarily entails” I mean that every time discrimination occurs, a pattern separation mechanism is at play, 100% of the time (or, >95% for us biologists 🙂 ). We need to know that when our car-park confusion is reduced, it is because of pattern separation. In order to rely on the conclusions of this logical form, it is not enough to assume that pattern separation could be a cause of our reduced car-park confusion.
My position is that discrimination does not entail pattern separation. In fact, it has not even been demonstrated that pattern separation (specifically in the dentate gyrus) even occurs at all!! This was the basis of my paper. Briefly put, my points were as follows:
1) It has not been experimentally demonstrated that pattern separation at the level of the cell population occurs.
2) Discrimination does not entail pattern separation since discrimination can be caused by other mechanisms (I discussed rate mapping, often used under the umbrella as “pattern separation,” as one potential mechanism. Cellular pattern separation in regions upstream of the dentate gyrus is another mechanism. Computational processes going on in the sensory systems is another).
A note on behavioral neuroscience. It has been critiqued that this is a matter of understanding how behavioral neuroscience “works,” but it is really not. Bussey correctly states that in behavioral neuroscience, we postulate “a putative process/construct/computation” and we “devise tasks to try to capture that function.” Importantly though, the examples given (and truthfully, the only valid examples I can think of) for the validity of this approach are for tasks that try to capture cognitive/behavioral functions. For example, we can have a task that tries to capture working memory, or attention. Similarly, our-touch screen, or radial arm maze, or context fear discrimination task tries to capture the behavioral phenomenon called “discrimination.” This type of workflow is completely valid, and is the foundation of behavioral neuroscience. I’d like to emphasize again that this workflow considers tasks that capture cognitive functions. Not tasks that capture underlying mechanisms. To do that we turn to molecular/cellular neuroscience. For example, you aren’t going to devise a behavioral task that tries to capture LTP, or synaptogenesis. You simply measure those things directly instead. The task can capture some cognitive function that depends on LTP – but you’d have to actually measure LTP in your experiment in order to make any conclusions about it. You wouldn’t perform the cognitive task and state that LTP occurs without measuring it (unless, of course, previous work established that the task necessarily entails LTP).
Thus, this workflow has its bounds. It falls apart is if we use the tasks to capture 1) unproven mechanisms of behavioral functions, and 2) mechanisms that are not necessarily entailed by the behavioral functions. As I stated above, we need to be quite certain that a cognitive phenomenon necessarily entails a particular mechanism (and that the mechanism even exists to begin with) if we are to make proper conclusions. So if, for example, a certain task necessarily entails LTP somewhere in an underlying circuit, then sure, we can say that the task “captures” this LTP mechanism. But this is trivial, and of no use. We aren’t doing behavioral neuroscience tasks to capture cellular mechanisms. The purpose of the workflow is to evaluate cognitive functions. This is how we measure cognitive functions. For the mechanisms, we have our other (direct) measurements.
Bussey says:
“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.”
Indeed, pattern separation putatively results in increasing discriminability. The task, however, tries to capture event discrimination. The task does not try to capture pattern separation. If transcription of protein-A putatively causes memory interference,why don’t we use a task that captures memory interference and then make conclusions on the transcription of protein-A without actually measuring protein-A transcription? According to the original computational work, pattern separation is a ubiquitous information processing phenomenon, so it is just as “widespread” as LTP, and surely isn’t unique to spatial discrimination. Pattern separation, and pattern convergence, happen everywhere, all the time. It is simply a cellular process that is begging for measurement.
So what is the “problem with pattern separation.” In my opinion, there are many:
1) Some pattern separation literature affirms the consequent (1. If pattern separation occurs, then the animal discriminates. 2. The animal discriminates. 3. Therefore, pattern separation occurs.)
2) Some literature that does not affirm the consequent assumes the truth of the premise that discrimination (and the tasks used to capture discrimination) necessarily entails a pattern separation mechanism (specifically in the dentate gyrus). We don’t know if other mechanisms are at play.
3) There is disagreement as to whether pattern separation at the level of the cell population even occurs. It’s wonderful as a computational idea, but it needs to be experimentally demonstrated.
4) No study has looked at cell population activity in the DG’s input and output and has causally established the role of the DG in pattern separation.
This is all just my humble opinion, and I appreciate all the effort put forth to the discussion. A lot of the I left out any discussion of defining pattern separation behaviorally. I think the problems of having multiple definitions of a single term are easy to spot.
10 thoughts on “Pattern Separation – What’s the problem?! Part 2”
I suppose I’ll lend the voice of a lowly rising second year PhD student! As a relative rookie to the concept of pattern separation and all the baggage it may or may not carry, perhaps it will offer a different sort of perspective.
I can most certainly appreciate the need for rigor in defining and applying terminology. In light of some recent work, it does seem that ‘pattern separation’ has taken on something of a buzz-word status, and that it’s being used in ways that deviate in varying degrees from its intended purpose. This seems most obviously problematic for purely behavioral studies that make strong and direct neurobiological or neurocomputational claims in the absence of…well…any neurobiological or neurocomputational data.
However, I tend to agree with Craig that avoidance of a term because of a potential for confusion is not the answer, and may actually be a hindrance to what otherwise amounts to perfectly good science. I think that a more reasonable approach is to take caveats for what they are, and recognize that certain approaches simply necessitate them. Perhaps ‘pattern separation’ is not the most kosher term to use in the absence of neurobiological correlates, but provided that such correlates are reasonably established in a given line of research, it’s acceptable (in my humble opinion, of course) provided that the authors make their intent in using the term as clear as possible. Simply put, we may gather the eagerly-awaited direct evidence of transformationally orthogonalized input as it passes from LEC/MEC to DG/CA3 to CA1, but we then have to make yet another series of computational leaps that bring this information to behavior. It seems that an overwhelming amount of data is consistent with the assumption that the DG is especially suited to undertake pattern separation (though certainly it occurs elsewhere in the brain to various extents, which I doubt anyone would deny), and we have a plethora of behavioral findings that are perfectly consistent with this. It will certainly be helpful to consider the respective inputs from the LEC and MEC to the DG, as has been done in several animal models, as well as the outputs from the DG. Certainly, we ought to exercise caution in terms of how strongly we make our claims in the absence of established black-and-white fact, but this applies to quite a large proportion of neuroscientific research. If the shoe fits, I see little harm in discussing things in terms of pattern separation (or pattern separation very likely having occurred) provided that we openly recognize the limitations of a given study and agree upon a set of reasonable standards.
There’s something to be said for avoiding operating over fallacious logical inferences. On the other hand, as several posts have pointed out, these fallacies are not necessarily so widespread (and don’t themselves necessitate that the conclusion was false). I do think that our field out to expect logical points to be made in publications. However, we also ought to bear in mind that there are certainly many of us who make every attempt to be as transparent as possible about our assumptions and what they mean in terms of the conclusions we draw, and temper them accordingly. As a (hopefully) up-and-coming student, I would really hate to poke my head up in the middle of a witch hunt =)
That’s my two-cents, albeit longer-winded than I’d intended. The discussions on here thus far have been very interesting and at times pretty eye-opening. What I would really find useful is, as I alluded above, some sort of agreement on a reasonable set of standards, be they experimental or terminological, that can get us all on the same page (or at least in the same paragraph).
I don’t disagree that car-parking is not exactly process-pure but I don’t think that was Tim’s intent from this example or anyone else’s for that matter. Just wanted to clarify that and defend the use of such an example as I think it exemplifies scenarios under which pattern separation might be a helpful thing!
And I will take this opportunity once again to voice support for trying to understand what exactly it is that MEC/LEC are doing in service of encoding experiences and how do their computations affect downstream DG pattern separation. I think this is a critical avenue to investigate (and not just because we’re doing it) and hope that there’s a confluence of data both from humans and animals and across several different approaches and techniques that begin to speak to this in the coming years. This will significantly improve our understanding of how the DG encodes conjunctive representations.
Sorry if this is backtracking a bit, but I’m still catching up on all the posts. Adam takes issue with Tim’s example of where pattern separation is necessary. I think this may be a bit misplaced. Tim’s definition and indeed the example he offers is one that I as well as many many others have used in the past to describe pattern separation. It is not simply a discrimination procedure. It is a discrimination under conditions of interference that must be minimized. I would argue that such a situation does require pattern separation (cf the definition I used on the front page for what I mean here). Rate remapping is but one type, sensory discrimination of similar may be another. I think that Tim and I tend of think of separation as a much more universal phenomenon that happens across systems all the way from sensory to higher cognitive but vary in the mechanism and the information content.
So I’m not sure Tim’s example is not a valid example of where pattern separation is needed. There is interference that needs to be minimized. Where it is minimized and what mechanism exactly is something we can ask, but the definition of PS doesn’t include either rate remapping or dentate gyrus in it, remember? As long as overlap in input exceeds overlap in output, we call it separation. I don’t see any possible scenario under which minimizing the car park confusion wouldn’t require this particular algorithm to some extent. May be there’s something I’m missing in the argument.
If Tim’s description said that car park confusion can only be reduced by having DG pattern separation, I might take issue with it, but I don’t think either Tim or anyone else can make that claim strongly absent data with a car-park confusion experiment!!
Again, too late in the evening so if I don’t make any sense, I apologize.
I don’t know if I necessarily take issue that pattern separation is needed for these tasks. I think that pattern separation, as a neural mechanism, definitely underlies memory interference tasks. But I also think it underlies all sorts of other behavior, and that it occurs in many regions throughout the brain. So my stance isn’t so much that memory interference doesn’t entail pattern separation, it is that we don’t know how much pattern separation in the DG, if it is occurring, is actually contributing to behavior in these tasks.
Also, as stated in another post, pattern separation in the DG is reducing the inputs from the MEC and LEC. We don’t know how memory interference tasks manifest in population activity in the MEC/LEC, and exactly what sort of information processing the DG is accomplishing by separating the incoming info. What kind of information does the MEC/LEC hold in these tasks, and how is this information transformed via separation? The theory is that it will manifest behaviorally as reduced interference, but I think there are so many unknown aspects (and known intricacies) of DG function that we can’t really isolate pattern separation in the DG as causing reduced memory interference. I think it also gets a little bit tricky when you consider the pattern completion mechanism occurring in the CA3 – how does the DG’s input to the CA3 influence pattern completion, and what is actually occurring to population activity in both regions during these tasks? If the DG ultimately separates population activity input caused by 2 similar stimuli, but the CA3 exhibits highly overlapped population activity due to pattern completion, then what is the DG doing, and how does this manifest behaviorally?
Also, kind of as an aside technicality, but I think the car-park example can also have some contributions from simple “forgetting” in addition to any potential effects of pattern separation…especially since the assumption for pattern separation would be that the person successfully remembers each spot in which he parked, and is able to parse them day by day. The actual situation would be a forgetting of the previous day’s car location, and selective memory of the present day’s car location. Thus, not strictly “pattern separation,” per se.
One quick point here that bears mentioning and was actually the reason Ray and I wrote our review. The DG does not engage in pattern completion as defined.
If the DG does not engage I pattern. Separation or one reason another, that is lack of pattern separation. Not pattern completion.
For pattern completion, it is likely that a prominent but not talked about nearly enough pathway from the perforate path directly to CA3 is sufficient to provide a cue to guide retrieval, and particularly important for providing the partial cues to guide pattern completion. This was an idea elegantly discussed by Alessandro Treves and Edmund Rolls in 1992 and has since been critical to their models.
Additionally, and this is important for how the DG does not always dominate and thus we only pattern separate, the CA3 pyramidal cells’ spike 2-4 ms earlier than DG granule cells to pp stimulation. Brian Derrick has shown this for both MPP and LPP Stimulation in vivo.
As such, the DG does not play a role in pattern completion. In fact, based on the models, it cannot play a role due to the competitive inhibitory network that has been proposed to underlie the actual performance of the pattern separation process.
Just my two cents on this topic since we need to be careful not to misdeed e pattern completion. Especially since we are trying to prevent mis defining pattern separation.
Craig,
I definitely see the value in using the term “behavioral pattern separation” – you are obviously right that discrimination has its baggage, and that a new term, properly defined, will help with the confusion (and help our pubmed searches). I think my main concerns stem from my citation analysis. In paper after paper I saw statements that claimed the existence of pattern separation and cited computational work, behavioral work, cellular work, etc. If we are to define “behavioral pattern separation” as a unique behavioral process (one that is distinct from whatever definition of “discrimination” that we use) then care must also be taken to not equate it with the underlying computational process. As you stated, behavioral pattern separation is consistent with computational pattern separation (it’s also consistent with a number of other cellular and molecular phenomena too, making me question why we even bother stating that it is consistent with the computational term).
I’m not sure I really agree that the “baggage” associated with “pattern separation” is equivalent, or lesser than that associated with “discrimination.” I think this is because discrimination, however defined, ultimately describes some sort of behavior. However, pattern separation defines a neurocomputational/cellular process. Thus, adding a modifier to pattern separation to turn it into a behavioral phenomenon is more confusing (to me at least) than simply trying to qualify the type of discrimination occurring.
” Yes, we hope that “behavioral pattern separation” is indexing actual pattern separation.”
If pattern separation is simply a computational tool used in networks throughout the brain, much like how LTP is a cellular “tool” used by neurons, then is it really valuable to always link “behavioral pattern separation” back to computational pattern separation? Why not keep them as completely different terms, with different meanings, at different levels of neuroscience? Computational pattern separation is definitely not DG specific – why not try to link behavioral pattern separation with computational pattern separation occurring in the sensory cortices? I vote to just keep the terms separate, describing separate phenomena.
Regarding Tim’s last point, I agree with Tim here that while one can call a task a pattern separation task using the “task taps into construct” paradigm, a term like “behavioral pattern separation” has a wholly different set of issues, which is that this “pattern separation” may or may not have anything to do with the neural computation. To me, these are orthogonal issues. One has to do with the logic of psychological experimentation and I think Tim gives some good examples for why this is not a problem. The other is really perhaps a semantic or definitional issue that does have the capacity to confuse if misused…
My approach here would be to call a task by some descriptor that’s task related, so a task that asks a subject to discriminate among different stimuli in memory is really a discrimination task that may or may not be sensitive to DG pattern separation (one can make the latter link with doing the neurobiology legwork), but the measure of behavior is simply a discrimination index. I would in theory be fine with calling it a pattern separation task (as long as it’s clearly defined) but I wouldn’t necessarily call the behavioral responses pattern separation behavior or refer to “behavioral pattern separation” for the same reasons Tim lists above.
>I state routinely that such behavior is consistent with pattern separation, not that it is pattern separation.
Yes. We don’t say task performance = the construct in any domain, be it working memory, pattern separation, or whatever. It’s a straw man.
>nobody has observed true pattern separation
Yes, just as no one has ‘observed’ working memory.
>“behavioral pattern separation”
Here I disagree, for the reasons in my initial post. Would you introduce the term “behavioral working memory”?
Tim
Adam,
We’re all well aware of logical fallacies and of problems of reverse inference. I will state here, as I have stated in every talk on this matter, that I cannot know that pattern separation has occurred in my behavioral tasks. I state routinely that such behavior is consistent with pattern separation, not that it is pattern separation.
I believe that your term discrimination is far more flawed than you appreciate. Yes, “pattern separation” has a history and a meaning, but so does “discrimination”. Is this one point or two in front of you? Is this line at the same angle as this other? Is this the same hue? These are all well-worn uses of the term “discrimination” and if we adopt it rampantly, we would need a modifier — “mnemonic discrimination” perhaps.
But, at this point, we’ve taken a term that has baggage and added a modifier to it to remove said baggage. How is this different that “behavioral pattern separation” where we’ve taken a term that has baggage and added a modifier to remove said baggage. At least in the BPS term, we’re using then modifier on a term that is inherently related to the area of research we’re working on — something that is not the case with mnemonic discrimination.
I would also posit that nobody has observed true pattern separation – the actual orthogonalization process. Take Vazdarjanova and Guzowski’s work — great stuff and I love it. But, their IEG expression is linked to the plasticity process, so this is a bit indirect (as behavior and fMRI are indirect – albeit to different degrees). Even the Leutgeb / Moser work where we have electrodes doing the recording of individual neurons — we’ve not observed the transformation process itself. Rather, we’ve observed the outcome – some result or effect of PS.
If we can make these arguments to any degree, then NOBODY should be able to use the term “pattern separation”. At which point, it’s equally useless as rampant overuse. So, do we all have modifiers?
This is the line of thinking that has led me to use “behavioral pattern separation”. I put a modifier on there for good reason. I cannot know the computation actually happened. I also always say my fMRI data are “consistent with a pattern separation process” having occurred. Again, I think I am not going down the logical fallacy path you are worried about by doing this.
But, adopting “discrimination” to me is not a better solution. If you type “discrimination” into PubMed, you get about 100k results. The vast majority of these have nothing to do with what we’re talking about. Thus, adopting that term is only going to increase confusion. If you type “pattern separation” in there, you get 187 results. “Behavioral pattern separation” has 4 (and yes, at least one isn’t mine!).
So long as we define our terms and so long as we are careful in our discussions and logic, I don’t see the problem. Yes, we hope that “behavioral pattern separation” is indexing actual pattern separation. This is why we attempt to link behavior in the task to things like activity in the DG/CA3, integrity of the input to the DG/CA3, functional connectivity in the EC-DG/CA3 pathway, etc. The tighter we can form this link, the more solid the relationship is between an example of behavioral pattern separation and actual neural pattern separation. (Note I said the “more solid” – this still has theoretical error bars on it.).
So, is “pattern separation” getting over-used? Sure – and we should make sure things get clarified and that we’re all careful. Should we ditch the term entirely unless we’re 100% certain it’s being used in the “pure” or “true” way? I think that’s throwing the proverbial baby out with the bath water and will do more harm than good.
Craig
I think Ray has an argument that is almost always glossed over in research into pattern separation (cf., Link ). The point of papers such as this is that the best way we can determine a behavioral output of pattern separation is by asking the participant or animal to make a series of discriminations wherein the interference among stimuli is parametrically modulated. I these cases, interference is used as a proxy for the similarity of input patterns that must be orthogonalized for efficient storage.
In our rat work, that means making the distance between a target and foil vary in distance across trials and plotting behavioral performance as a function of interference (this was the point that Ray and I may have clumsily made in plotting pattern separation functions across attributes in our latest review on the topic). For humans, Brock, Craig and Mike among others have done the same thing by parameterizing the stimuli and showing that they can select objects that can be scaled by similarity I such a way that is appears at face value very pattern separation-y.
I think a focus on the specifics of task design will actually help overcome the logical fallacies Adam is pointing out. The simple fact is that we DO NOT know whether a participant is engaging a pattern separation process to perform a task. What we can do is to develop paradigms that allow us to specifically probe processes that either directly reflect pattern separation or rely very highly upon pattern separation to optimize task performance. As I stated in a comment to Tim’s post, it is definitely more intellectually honest to say “we see that performance on this task decreased as a function of increased interference, potentially reflecting an underlying pattern separation process” rather than “they showed intact pattern separation on this task”.
What we must avoid, and the main impetus for Ray and me writing the review in 2012/2013 was that it is becoming too easy for labs to say bad watermaze performance =bad memory=pattern separation deficit. I think the real goal that we can undertake as a group is to develop a set of clear definitions that can be used by molecular biologists and biological psychologists using animals models to study dentate gyrus function (or any other brain area that engages in pattern separation-like processes).
I have no idea what the solution is, but I think we can make a good start with these discussions.
Comments are closed.