There is a cohort of computational models of the hippocampus and its role in memory that have similar properties. These models stress the role of the hippocampus in relatively rapid, arbitrary associative learning and use pattern separation to alleviate potential high levels of interference. These models have been tested and refined based on rodent recording and imaging work in several labs. By and large, there has been support for the general notion that the hippocampus, and the dentate gyrus in particular, is reducing the overlap in representations of similar information — aka, it is exhibiting pattern separation. While witnessing the transformation itself (so-as to satisfy the pure definition of pattern separation) is still a work in progress, the data have, by and large, supported these models and have certainly not refuted their general principles.

In my opinion, these models and these principles gain immense value if they can be extended to humans and to human behavior. Since 2006, my lab has been working to see if we could form the needed links. Any such links would, by necessity, be indirect. I know of no behavioral measure that has a direct readout of an underlying neural function of the CNS. Likewise, neuroimaging in humans is not a direct readout of neural activity. These preclude any direct link being formed and preclude any strong reverse inferences from being made. That is why I strive to ensure that my claims show “signals consistent with pattern separation” rather than being observations of pattern separation. In my mind, this area is one in which we can profit greatly from the convergence across techniques and levels of analyses.

On this site and elsewhere, there has been considerable confusion and debate about the term “behavioral pattern separation”. Some feel the term implies a 1:1 mapping between the computational form of pattern separation and behavioral performance. Others feel it refers to a non-linear transfer function between input similarity and behavior. Thus, any categorization or classification task would fit such a definition, for example.

Neither of these are what I have meant by the term. The term was originally intended to describe a task that was developed in the lab that had the goal of behaviorally assessing the integrity of pattern separation in the hippocampus. Of course it would never be a 1:1 mapping making reverse inferences problematic (other problems in the system might lead to an alteration of the behavior). But, the goal was to try to design a task that, were these models generally correct, would rely upon the computations being performed by the hippocampus (and the dentate in particular) and would provide some measurable signal. The task was not originally called a “behavioral pattern separation” task but was called a task that “taxes pattern separation” and later a “mnemonic similarity task”. Only in the one most recent paper have we used the “BPS” moniker.

In our initial report (Kirwan & Stark, 2007) we said:
“That is not to say that the predictions of the computational models cannot be tested using fMRI, however. Here we present two experiments that test a fundamental hypothesis that the hippocampus possesses the strongest ability to perform pattern separation (e.g., via orthogonalization in the dentate gyrus). We do this using a variant of the continuous recognition task that taxes pattern separation. The behavioral consequence of pattern separation is the ability to mnemonically distinguish between two similar stimuli. In order to behaviorally tax pattern separation, one should be able to present subjects with two similar stimuli (e.g., as a target and a similar lure in a continuous recognition memory paradigm) and determine if the subject can distinguish between them by correctly rejecting the lures as being new and only similar to previously presented items. The ability to reject similar lures critically depends on pattern separation.”

I should note that we we went on to note that, “This approach, while allowing us to place demands on pattern separation processes, still does not allow us to cleanly isolate computational pattern separation.” Since the outset, and I believe in every time since, we have stressed the notion that the task is designed to tax such a system or rely on the operation of such a system. But, it’s never been put forth as some perfect measure of computational pattern separation.

In subsequent work, we have tied this performance to age-related changes in the input to the DG and CA3, to functional signals in the DG/CA3, to modulation of the system via NE (which has robust input to the DG), etc. We’ve shown different input-output transfer functions and done a good amount of work to “neurobiologically validate” the task and its link to the hippocampus and the dentate in particular. The behavioral task has been astoundingly reliable and it’s been useful not only to our lab but also to others.

Yet, can we or should we call it “behavioral pattern separation”? There are issues here. There may very well be other computations (not part of the current modeling zeitgeist) in the hippocampus that let it resist interference much in the way pattern separation does but that don’t employ this computation. Our results might fit such models equally well, even though no pattern separation was involved. We are indirect in our measures and naming the task this may be a bridge too far. At the very least, it is clear that it has caused confusion. I certainly do not want to suggest that any task that has any kind of discrimination component to it is “pattern separation” and that, by virtue of this, is tapping into the dentate. Holding up my hand for someone and asking “how many fingers am I holding up?” is not what I had in mind and while there may-well be merit in understanding commonalities in any discrimination paradigm or any non-linear transfer function, that’s not what I was intending when I proposed the term.

What’s in a name may not be hand, nor foot, nor any other part of a man, but it clearly can have some large consequences. As a result, I am changing the name of this task from the BPS back to its prior name, the “Mnemonic Similarity Task”. Our metric in this task will not be a BPS score, but a lure discrimination index (LDI) or simply the formula.

While any direct link to the computational notion of pattern separation is impossible, I still believe there is incredible value in attempting such a link. Yes, it will be noisy and yes, reverse inferences are problematic at the very best. But what value is a model of memory that is never applied to behavior (and, ideally, to human behavior)? To bring in another quote, we do not try this because it is easy, but because it is hard. To me, that’s where the challenge is, where the fun is, and where the reward is. We’ve taken a number of steps to form this link, but more are needed. Heck, more are needed to observe pattern separation in action in even a rodent hippocampus. It’s hard, but people are working on it and making progress.

This area is, in my opinion, an ideal place where we can try to fulfill the promise of integrative neuroscience or cognitive neuroscience. We have similar ideas and people working at every level that could provide a compelling set of answers to how this form of memory works and why the system has been designed in such a way. At each level, the links won’t be perfect, but that will hopefully let us refine our thinking. So, while I will refrain from the term “behavioral pattern separation” I’m not backing away from the ideas that led to the term – that we can, and should, try to form these links.

Craig

What’s in a name?