Please post your comments on multiple systems models, and more generally on the relationship between memory and concept learning, here (related to class 10/7).
Please post your comments on multiple systems models, and more generally on the relationship between memory and concept learning, here (related to class 10/7).
This comment is aimed to clarify a little bit about Wong et al. mean by “concatenative grammar” and “analogical grammar”, and to help us figure out what might be going on in their results. They state in the introduction that “concatenative grammar learning is associated with the fronto-striatal system and procedural memory, while analogical grammar learning is at least partially dissociable from concatenative grammar learning and is associated with declarative memory.” One thing that may be a bit confusing is that a single language is being learned in the experiment – we often associate one “grammar” with one “language”. Because the analogical and concatenative results come from the same experiment, it’s also little hard to figure out what’s going on, especially since both patterns involve learning of the same prefixes and suffixes. The prefix is always of form [ki]. When it is added to a stem that contains /e/, the /e/ changes to [i]. The suffix alternates between [el] and [il], depending on the vowel of the stem. The tricky thing is that the suffix is [el] following an /e/ stem, even when it has the [ki] prefix, and it’s vowel has changed to [i]. You can see illustrations in their figure 3. As you can see there too, they are using the term “concatenative” for items in which there is no change to the vowel, and “analogical” for cases where the vowel changes (analogical usually has a different sense – another confusing thing). I can’t tell yet whether they want analogical to refer just to cases like what I labeled “tricky” – what phonologists call opacity, or whether this means any case of alternation. Why these should map to declarative and procedural memory I don’t yet quite understand – perhaps the complicated blocking procedure they used would play a role (see Language Learning), since they would need to learn across blocks that the final vowel of ki-pish-el is the way it is because of the form in the singular or plural. (thanks to Claire for discussion)
The distinction between rule-based and information-integration tasks is unsatisfyingly vague. Ashby and Maddox explicitly say that the ends of the spectrum may fit into either category, but where the boundary is between the two ends, or at which level of complexity does a task cease to be rule-based and become information-integration, is less clear. The distinction between verbalizable and non-verbalizable is also not particularly helpful because the verbalization need not be on the part of the learner; there are apparently “no assumptions about how people learn these different category structures in any particular application (pg. 154).†The distinction is empirical, based on dissociations.
However, as Newell and Dunn (2008) and others have pointed out, dissociations rely on the flawed logic that an underlying psychological process—or crucial brain area—affects performance on one task positively could not affect performance on another task negatively. State-trace analysis has been one method to bypass flawed dissociation logic, but is there another way to actually empirically distinguish between the two types of tasks?
Also, the conclusion that rule-based tasks must be simple enough that amensiacs with medial temporal lobe damage can use working memory to do them is a little confusing to me. Couldn’t the task be simple enough that amnesiacs could use procedural memory to do them? I would think that a simple task would be learnable, whether or not the rule is verbalizable, and I feel like procedural memory is a lower level cognition that could explain a simple categorization task without resorting to higher level working memory. This would be more parsimonious because we already know that amnesiacs’ procedural memory remains intact, and indeed Ashby and Maddox discuss this for the information-integration tasks (pg. 160).
In the Ashby & Maddox (2005) review, they mention executive attention as more of a factor in rule-based category learning than information-integration tasks. They cited a dual-task study (Waldron & Ashby, 2001) demonstrating impaired performance on rule-based learning, but not information-integration learning, when the other task involved working memory and executive attention. This seems to make a pretty strong case for the existence of more than one type of category learning, so I wondered why they didn’t look at any populations with executive function impairment (such as adults with ADHD). I couldn’t find anything in a lit search on that topic either; does anyone know why this hasn’t really been investigated? I’d be interested to see if this population showed compensatory mechanisms in rule-based category learning or more deficits in that area than on other category-learning tasks.
Monica
I’m trying to integrate your comment and the Ashby and Maddox paper with the Lewandowsky paper that I presented. Do you know if people with ADHD also generally have a lower working memory capacity? I would think so, given what I know about both, but my knowledge of special populations is pretty limited.
I am curious about the issue of the “easy verbalization”, which is necessary for the rule based learning. I am wondering how much this trait is language/culture/individual dependent. Assume that we have an experiment in which the participants need to lean that category A is the green + checkered + triangle objects. Suppose there is a language in which there is a concept for ‘green and checkered” property and it is encoded in one word. Is there evidence that such people would have to allocate smaller cognitive resources for the task, since they will have to attend to a conjunction of just two properties (“green + checkered” and triangle) as opposed to the English speakers (green + checkered + triangle)?
Regarding Wong et al : I admit that I didn’t understand a lot in their article, but just about their hypothesis: “More specifically, previous research [10] and our current experiment (see below) confirmed that concatenative grammar learning is associated with the fronto-striatal system and procedural memory, while analogical grammar learning is at least partially dissociable from concatenative grammar learning and is associated with declarative memory.
Do they mean that native speakers of an English – type language and of a Semitic language will, as a result of their specific language exposure, have different memory capacities?
I am aware that they are talking about L2 acquisition.
Ashby & Maddox’s article includes a brief discussion of separable vs. integral features. As I have heard discussions of phonological features under this distinction, they have been characterized as integral (I believe that Nosofsky 1992? talks about them this way?). But it strikes me that some distinctive features might be more separable than others, in the sense that listeners could monitor better for them independent of the actual identity of phones carrying them (perhaps [strident] or [nasal]?). Could it be the case that some phonological features, just because of the nature of their physical correlates, lend themselves to the use of more learning types than others?
Tina: Thanks for the pointer to Newell and Dunn (2008) – Ben Newell’s web page turned out to be a goldmine! I’ve posted a recent critical review of multiple systems models to our syllabus page as a good thing to read if you are interested in this topic. Now I know why Kurtz et al. were so reticent about going for a multiple systems explantation for their result!
Everyone: I’m focusing for the moment on trying to get my head around some of this massive literature, and thinking about the connections to the Shepard II vs. IV stuff we’ve been talking about in class. More soon.
Ashby and Maddox (2005) Summary:
Ashby and Maddox (2005) present a case for multiple memory systems approach to category learning tasks. In particular, rule-based tasks, which require verbalizable reasoning, are thought to depend mostly on working memory and executive attention. Information-integration tasks, which are not verbalizable and are sensitive to the type of feedback, are thought to depend mostly on procedural learning. Prototype distortion tasks are thought to depend mostly on perceptual learning. Weather prediction tasks, which are probabilistic, may be completed using the different strategies. The evidence for these distinctions come from behavioral data as well as neuroimaging data.
Given my skepticism of relying heavily on empirical dissociations to distinguish cognitive processes as well as the vague boundaries between some of the tasks, I’m not sure how convinced I am of the multiple systems idea for category learning. Newell, Dunn, & Kalish (2010) provided further evidence against the multiple systems approach using state-trace analysis on ERP data.
Summary for Ullman (2012):
Ullman (2012) summarizes the Declarative/Procedural model of memory and language. The neural substrate of the declarative memory system (knowledge about facts and events) is in the hippocampus, temporal lobes, and some areas of frontal cortex. In contrast, the procedural memory system (implicit memory for skills involving action, such as motor skills) is based in the basal ganglia and connected frontal cortex. This model posits that the declarative memory system underlies lexical information and all such “idiosyncratic information” in language, and the procedural memory system underlies the parts of language that involve manipulation using rules (e.g., morphology, syntax). These predictions concern the speaker’s L1; the model predicts that an L2 language-learner will rely more heavily on declarative memory than procedural memory early on in acquisition, in contrast to the reverse prediction for L1 acquisition. This reliance is predicted to shift with more profiency to procedural memory for L2 learners.
The author presents evidence from the ERP literature supporting this model. He suggests that because the N400 is localized to bilateral, central and posterior locations and involved in lexical-semantic processing, this reflects overlap with the declarative memory system. Similarly, the P600 and LAN involve morphosyntactic processing and tend to be localized to central and parietal areas, and are suggested to be related to procedural memory. There does not appear to be a difference in N400 findings for L2 speakers, while the LAN tends to be absent for lower-proficiency L2 speakers, supporting the assertions of the D/P model about acquisition. Similar results were found for artificial language learning.