Category Archives: Lab Meetings

Cognitive Interference and Category Learning: A Tale of Two Systems

Joshua Hatherley, a PhD student in the Minda Cognitive Science lab, is presenting some research from his dissertation as a poster at the 2019 Psychonomics Meeting in Montreal. The poster presentation is scheduled for Friday, November 15, 2019 from 12:00 PM – 1:30 PM in Room 517B. You can get a copy of the poster link here.

The Experiment

Josh investigated the relationship between working memory, procedural memory, and category learning. The COVIS model assumes that category learning is regulated by an explicit system which uses working memory and an implicit system which uses procedural memory. Research by Minda et al. (2008) used a co-articulation task to disrupt disjunctive-rule category learning and Miles and Minda (2011) used a switching task to disrupt rule-based category learning. However, Minda et al. (2008) failed to find a link between performance on single-dimensional category sets and working memory. Miles and Minda (2011) used a demanding task that interfered with verbal working memory, visual working memory, and executive functions but did not specify if one or all of these processes caused the disruption. The present study addressed these limitations by asking people to learn either a rule-defined (RD) or a non-rule-defined (Information Integration) category set (see Figure 1) composed of 80 Gabor patch  images, 40 in each category, that varied in either the orientation of the image, or the frequency of the lines in the image.

While completing this primary categorization task, participants were asked to also perform one of three concurrent tasks that interfered with either verbal working memory, procedural memory, or none of the processes (see Figure 2).  In the first concurrent task, participants only learned the category set. This was done in order to be able to compare concurrent task performance to some kind of baseline. In the second concurrent task, participants were asked to speak aloud letters as they appear directly underneath the categorization stimuli. As reading and speaking a list of letters as they appear on the screen should have used a verbal working memory system, it was expected that this additional task would divert resources away from working memory and interfere with learning rule-defined category sets. In the third concurrent task, participants were assigned to complete a motor concurrent task. In it, participants were asked to tap the table with their non-dominant hand whenever they see a colon appear on the screen. As motor function is thought to be a process of procedural memory, adding the additional burden of performing a tapping task should slow down participants ability to learn non-rule-defined categories.

Results

Results indicated that both the verbal working memory task and the procedural memory task impaired the learning of rule-defined categories but had no effect on the learning of information integration categories (see Figure 3).

Information integration category learning did not seem to be affected by either the verbal or the tapping concurrent task. Although we did not  predict an effect of the verbal concurrent task, we did predict an effect of the tapping task. We also see that rule-defined category learning appears to be slowed by both the verbal concurrent task and the tapping concurrent task. While we did expect to see an impact from the verbal concurrent task on rule-defined category learning, we did not expect to see it from the tapping concurrent task – which was thought to only impact procedural memory. 

What do these results mean? For one thing, learning to perform perceptual classifications is difficult. It takes practice. But people can learn to do it fairly well, even for stimuli like these that are hard to describe. But these results also show that some classifications, the rule-defined ones,  are made even more difficult when people are asked to do two things at once. We think that the cognitive resources that are needed to speak the letters or to switch between tasks are also being used for learning these categories. Other classification, the information-integration categories, do not seem to suffer from the same kind of interference. Do these results suggest that there are two cognitive systems that learn new categories? Perhaps. The best way to know for sure is to look more closely at some of the cognitive and neural substrates involved in learning. That’s what our lab is planning to do next. 

This work was supported by NSERC, Western University, and BrainsCAN.

For full analysis, see our rpubs page.


People Are More Likely to Use Classification Rules When Features Are Easy to Describe Verbally

Bailey Brashears, a PhD in the Minda Cognitive Science lab, is presenting some research from her Master’s thesis as a poster at the 2019 Psychonomics Meeting in Montreal. The poster presentation is scheduled for the Friday Nov 15 poster session at noon. You can get a copy of the poster here

The Experiment

Bailey investigated the effects of feature verbalizablity on the acquisition of novel concepts. Participants in the experiment learned a category set that could be acquired by either a perfectly predictive criterial attribute rule or an overall family-resemblance strategy. Half of the participants learned this set with features that were easy to name and describe verbally and the other half learned this set with features that were not easy to name and describe verbally. In this case, easy to name meant features that corresponded to focal colours, nameable shapes, or countable shapes. Features that were not easy to name were equally diagnostic of category membership, but did not correspond to focal colours or nameable shapes.  

Some examples of the stimuli we used. Each pair shown are opposite prototypes

Results

After learning, participants were transferred to a set of new exemplars that included stimuli designed to distinguish between rule strategies and family-resemblance strategies. We found that it took participants about the same number of trials to learn the category sets regardless of the feature type learned. However, participants’ accuracy on previous items in the testing phase was higher for participants who learned the stimulus with easy to verbalize features; while participants were able to learn the categories in either condition, they retained these categories with better accuracy when the features were easy to verbalize. 

The majority of the subjects in the easily verbalizable condition were fit best by the criterial attribute (CA) rule model.

We also analyzed each learner’s performance with a set of computational models. Each model in the set assumed that performance was driven by one (and only one) of the available strategies (a rule, family resemblance, guessing, etc) and we examined which of the model best fit the observed performances. We found that people who learned the categories with easier to name features were more likely to classify new stimuli in accordance with a rule-based strategy. People who learned the categories with difficult to name features showed evidence of both rule use and family resemblance responding and no clear preference for either strategy. The figure above shows the models’ fit (essentially how close the model is to the observed data) for each subject by condition.

Implications

What do these results mean? For one thing, these results suggest that when people can rely on language to label features and to name things, they do. The more available the names are, the more likely people are to use rules that correspond to those features. This points to the primacy of language as a means to consolidate information. This language-based rule system might be a default way to acquire new concepts. It also means that there are other ways to learn concepts beyond using language, however. People who learned the categories that had features that were hard to describe verbally often learned the overall family resemblance, suggesting a tendency to learn exemplars. 

Bailey is working on a second study to replicate these results and plans to develop other ways to explore the role of language in acquiring new concepts. As well, she’s working on designing gaming environments that can explore the incidental acquisition of concepts. Stay tuned for more!


This work was supported by NSERC, Western University, and BrainsCAN.

Measuring EEG during a category learning task

Humans learn to classify and categorize things all the time. It’s a fundamental aspect of cognition. Sometimes, categories can be learned in more than one way and there may be some things which seem to be exceptions to the classification rule. How does the brain react to individual stimuli during the learning process and in particular, how does the brain deal with learning these exceptions?

A new paper from our lab, published at Cognitive, Affective, and Behavioral Neuroscience, tries to answer this question. This work was carried out by a former lab member and PhD student, Rachel Rabi (who is now Banting postdoc working with Nicole Anderson and Lyne Hasher at the Rotman Research Institute in Toronto) along with Marc Joanisse, and lab members Toka Zhu and Paul Minda.

Learning a Conjunctive Rule

In the study, participants CJB1_28learned to classify shapes known as Gabor Patches (seen to the right), which are fairly common in visual perception research. Two features varied on these stimuli: the orientation of the alternating light and dark bands and the spatial frequency of the light and dark bands. We created a whole set of these and the figure below shows how these stimuli were arranged according to variance along the Orientation and Frequency dimensions. The optimal categorization strategy for this set was the conjunctive rule that combined frequency and orientation. However, it was also possible to learn this category less than perfectly with one of two sub-optimal, single-dimensional rules that were easier to acquire. In the example below, if a person only used frequency, it would allow them to classify all the B items, along with the items in the A1 and A3 regions, but they would misclassify the items in the A2 region because just using frequency does not distinguish between those two. In this case, participant would make errors on those stimuli until they transitioned to the optimal rule.

Untitled

Our goal was to understand the cognitive processes behind the transition from a simple, suboptimal rule to a more complex optimal rule. We reasoned that this mirrors many real-world scenarios in which people learn a quick and easy rule that’s usually correct but with time and practice can use a more complicated rule that’s almost always correct.

Measuring EEG

Participants learned to classify the items in the set by seeing a single visual stimulus on a computer screen, making a decision as to whether it belonged in one category or the other (A or B), and then receiving feedback on whether or not they made the right decision. This feedback-driven learning went on for several hundred trials. During the learning phase, we also measured electrical activity in the brain using EEG. The EEG wave forms allowed us to see what was going in the brain at the moment that the participants saw the stimulus, made a decision, and received feedback.

We tested several hypotheses, and one hypothesis was that even after people learned to classify all the items correctly and had learned the conjunctive rule, we should be able to detect a difference between the stimuli that were able to be classified with only the simple rule (“easy” items) and those that could only be classified with the full, conjunctive rule (“difficult” items). We predicted that we’d find a difference in the EEG wave forms between these items, even though the behavioural response was the same.

Processing of “Exceptions”

We used a series of computational models to determine which strategy each person was using (details in the paper) and, as predicted, we found that people transitioned from a single-feature rule to the conjunctive rule. Then we looked at the EEG wave forms for participants who were using that conjunctive rule and making correct decisions. What we found as that even though they classified all of the stimuli correctly, there was a difference in brain activity between the “easy” stimuli and the “hard” stimuli. The difference can be seem below and it shows higher amplitude for the difficult stimuli, roughly 400-600 msec after as image is seen. We interpreted this difference as evidence for the additional memory processing that was happening when participants saw and classified the difficult stimuli.

It seems like even though you can make a correct decision on all of these items, your brain reacts differently to the difficult stimuli. Your brain seems to know that these items were, for a time, exceptions to a simple rule.

Untitled 2

We’re working on some new projects to understand more about how the brain reacts when learning about categories that have exceptional items. In addition, this was our first paper that combined the modelling work (which we’ve done a lot of) with EEG ( which we only done a little of). We plan to keep working with this technique. The modelling tells us what strategy participants most likely used and the EEG tells us how the brain was processed these stimuli. Together, it gives us a more complete picture of something that most of us just take for granted: making a quick classification decisions.

Find out More

You can see more about this work by reading the paper at the journal or the preprint from the preprint server.  If you really want to dig unto what we’be done, you can see all of the stimuli that we used, the individual data from out study and the code for our analyses and computation models at my Open Science Framework Page.

 

A Curated Reading List

This was originally posted on my personal blog, I’m reblogging here as well, since this is basic cognitive psych

The Modern Scientist

Fact: I do not read enough of the literature any more. I don’t really read anything. I read manuscripts that I am reviewing, but that’s not really sufficient to stay abreast of the field. I assign readings for classes, to grad students, and trainees and we may discuss current trends. This is great for lab, but for me the effect is something like me saying to my lab “read this and tell me what happened”. And I read twitter.

But I always have a list of things I want to read. What better way to work through these papers than to blog about them, right?

So this the first instalment of “Paul’s Curated Reading List”. I’m going to focus on cognitive science approaches to categorization and classification behaviour. That is my primary field, and the one I most want to stay abreast of. In each instalment, I’ll pick a…

View original post 1,262 more words

The Exemplar Model

We’re working to document and share our computational modelling code on Open Science Framework. One of the most widely used models is Rob Nosofsky’s Exemplar model known as the Generalized Context Model. This project can be found here, and is the result of Amanda Lien’s NSERC summer project in 2017. (Fun aside: I programmed a version of this model in 1998 using Turbo Pascal 7.0, later updated to REALBasic/Xojo, then a slow R script with loops, and finally with A. Lien’s help, an R version with matrices that minimizes running time. The underlying math is the same.)

This notebook describes the formulation of an exemplar model of categorization, the Generalized Context Model. Full specification of the model itself can be found elsewhere (Nosofsky 1984Minda & Smith 2001) This model has been used in cognitive psychological research to make predictions about how participants will learn to classify objects and belonging to one or more category. The primary assumption of the exemplar model is that categories are represented in the mind by stored exemplar traces, rather that rules or prototypes.

The document describes the development and use of an R script that reads in a text file of classification probabilities (usually obtained from behavioural testing), a text file that corresponds to the stimuli in the experiment, and a text file that corresponds to the exemplars of each category. The model then uses a hill-climbing algorithm to adjust the parameters and minimize the fitting error. The model will report the best-fitting parameters, the fit index, and the prediction of the model.

 

Canadian Scientists Organizing Boycott Against US Conferences

An article posted on Motherboard discusses the boycott of several Canadian scientists against attending US conferences as a form of protest against the Trump administration’s ban on travel from seven Muslim-majority countries. Western’s own neuroscientist Dr. Owen has offered to compensate the cancellation fees of others from Western’s Brain and Mind Institute and pay for international scientists to come to the university to present their research. Dr. Minda and the Categorization Lab are also considering joining the boycott, but a final decision has not yet been made.

Explaining the causal links between illness management and symptom reduction

Exciting new research published in Patient Education and Counseling was led by Karen Zhang, a recent grad from the Categorization Lab. Zhang and her fellow researchers revealed that providing explanation for why illness management is effective for reducing symptomatology can help improve the knowledge and application of health information for younger individuals. In contrast, reducing verbal demands of patient education material may help older adults learn new health information better.