6.804/9.66/9.660  Computational Cog Sci

Hey everyone,

Because of the additional length of PS3, we've decided to extend the deadline to Wednesday at midnight. I will be holding extra office hours tomorrow (Mon) from 4:30-5:30pm in 46-4053 and recitation will be focused on the problem set. I'll also have office hours on Tuesday after class, from 5-6pm. Please let me know if you need help but you can't make these times. Also, feel free to email me with any questions.

Best, Chris

Hi everyone,

Problem set 3 has been posted. It will be due on December 3 at midnight. The problem set is pretty long, so try to start early. Good luck!

Best, Chris

Hi everyone,

For the most part, great job on Problem Set 1! I hope it was worthwhile.

I've already gotten a flood of questions about this, so here are the stats:
Total points: 64 (20 for #1, 20 for #2, 24 for #3)
Mean: 58.55
Stdev: 6.81
Histogram: web.mit.edu/clbaker/www/pset1-hist.jpg

Most of you did really well, but if you scored less than 45 points, please contact one of the TAs. (We might also contact you). This is not out of concern for your grade so much as us wanting to make sure that you are learning the material.

I also have some general comments for each of the problems.

Problem 1:
-Vectorize code. I put this comment on a lot of your problem sets. What this means is that your code could have been simplified greatly by using Matlab's built-in functions that operate on vectors. Many of you relied heavily on "for" loops, like you would in C or Java, when instead you could have used the element-wise operators Matlab provides much more easily.
-Many students were skeptical of how well the models predicted their subjects' data. For the most part, everyone's model fit people's data very well. Predictions of computational models will never fit human judgments perfectly, because there are many factors we choose not to control for, and many aspects of cognition that we choose not to model in any given experiment.
-Although we asked you to speculate about the difference between cover story conditions, it was really difficult to draw conclusions about the effects of the cover stories from so few subjects. A related point is that it doesn't make sense to fuss with parameters to fit every nuance of your data when you only have noisy ratings from a few subjects.
-Finally, the best cover stories I read for this problem were the ones that were most vivid. These cover stories invoked strong imagery and gave subjects a sense of reality in the task. Anecdotally, these cover stories seemed to have the largest effect across all students' experiments.

Problem 2:
-Again, vectorize you code. Avoid for loops because they are slow in Matlab and often lead to complicated code with bugs.
-Many responses to this problem criticized the restricted hypothesis space of mathematical and interval concepts. Two things to note are that the actual paper used a broader range of hypotheses, and that the hypothesis space is not limited to a fixed set in any way. However, a deeper criticism is that the hypothesis space must be generated beforehand. A great project idea is to model how the hypothesis space can be generated using hierarchical Bayesian models. Something to note though, is that mathematical hypotheses are to some extent determined socially, pragmatically, or environmentally. So in this sense, it's actually ok to make a priori assumptions about the hypotheses that people might entertain (in my opinion).

Problem 3:
-For the numerical integration in compute_posterior, there were two main classes of responses, one that was technically more correct, but both of which we accepted. The way we had in mind was to consider postvals to be a discrete approximation to the continuous density. So sum(postvals)=1. Many of you treated postvals as a sample of delta functions  from the continuous density, which technically didn't work with Vikash's implementation of find_median(). Both implementations got full credit though.


General:
The easiest problem sets to grade were in PDF format, with responses and figures and code interleaved. Having the code and figures in a separate appendix made grading a lot more difficult because we had to skip back and forth.

Ok, that's it. Great job! Good luck on problem set 2!

Best, Chris

Hi everyone,

For tonight's recitation, bring questions about problem set 2 and recent lectures. It's not necessary to bring laptops unless you want to work on the problem set or have specific BUGS questions. Although the starter code for problem 3 isn't done, I can still answer questions or give an overview of Gibbs sampling.

Best, Chris

Hi everyone,

The writeup for problem set 2 has been posted. The starter code for problem 3 will be posted later tonight.

Problem set 1 is currently being graded, and will be finished by Monday.

Best, Chris