Apparently UCB doesn't have many campus maps, so I spent my first two hours on campus hauling my personal library in the exact opposite direction of my dorm. This was surprisingly not an enjoyable experience. By the time I got to my room, it was about one o'clock, and I finally had a chance to decompress after a day of exams, transit, and heavy lifting.
My first day in Boulder was brilliantly sunny, something I was prepared for, given my friend had told me it would be "sunny everyday." (Granted, he also told me everybody here would be attractive. Absolutely everyone.) But for any of you who've been in Boulder these past several days, it's been raining, and the temperature has dipped into the thirties. Bizarre. It seems like the apocalypse (Or are they calling it the rapture? I'm just not up to date on these things.) might actually come on the 21st. In the event that does happen, I would guess that means these journals would cease to be mandatory. One can hope.
As for the work I've been doing at the lab, I have been constructing what are called "exemplars" or "novel objects" for experiments to be held during the summer. The gist is that children's word acquisition must be located, and in order to do that, we must construct highly fabricated, pipe-cleaner-type contraptions so we can be sure, they have never seen these things before. They look odd, really. Perhaps I'll post a picture, once I get my hands on a camera.
Unsurprisingly, I have also been reading. Reading, reading, reading. Neural networks! Hill-climbing! Mutual exclusivity! It's all fascinating stuff, but it's still a lot. To my pleasure, I've found Mendeley to organize all the papers that I've googled and that Eliana has given me. Highly recommended, young DREU-ids. See:
On Wednesday, I went over to the farmers' market. I explored. I ate. I drank tea at the Dushanbe Teahouse. I'm aware that I got ripped off, but the teahouse is beautiful enough on its own that I didn't notice the lack of seasoning. The tea and decor are lovely.
How pretty.
Back to Top
It's gotten out of hand, Nature. Out of hand.
My work for this week mainly consisted of building this website, doing more reading, enduring the rain, walking through the rain, staying dry, preparing experiment materials, and getting a dyslexia model to work in Emergent.
Back to Top
On Monday, I thought it would take a single day to fully translate the MCDI English into Hindi. The test is given to young children in order to assess their "mastery" of the given language. I just finished the translation today, meaning that it took me a week. In retrospect, what was I thinking. Although I have taken Hindi for the past four semesters, I feel like an idiot it's a humbling experience to realize that the most illiterate child in India knows how to say 'butterfly,' whereas you do not. I'm trying to think of it now, and it's failing me. For shame.
Granted, in addition to translating, I did have to input the number of syllables (oh, shwa deletion), stress, diphthongs, and conjuncts. Whatever. Excuses.
The reason why I'm doing this at all is that I thought it would be a great extension on my original modeling project - simulate how children learn to read bilingually - if I added a Hindi-Urdu component. Since India has a history of bilingualism, it's a more pressing issue to understand how to best teach children to read and write in English and Hindi. Education policies are still being revised, and the three language formula (children learn Hindi, English, and a South Indian language) is still contested, despite being installed in 1968. There's an obvious disadvantage to children who don't live in states where Hindi is widely spoken. It's a similar situation to the impetus for this modeling project: When Spanish-speaking children enter American schools, they must learn to read and speak in English and predictably fall behind their peers, who already speak English at home. If this Hindi component works out, I'll be able to do follow-up research next fall, when I'll be in Hyderabad (Andhra Pradesh, where Telegu is predominantly spoken). Ahhh, full circle.
It's also sunny! Huzzah.
Back to Top
The week had some windy moments...
But I got to go to the Boulder Creek Festival over the weekend. Ten thousand people went on Saturday, which was when I was there. Absolutely insane. Growing up in a Jersey suburb, I find it strange (but lovely) that amidst all this suburbia, Boulderites have fun and in large numbers, too. Be more miserable, people. Have you not seen Garden State?
I also got the chance to go to the Ultimate Frisbee National championships. It's where old college hierarchies go to die. Carleton pummeled Harvard. Cornell sucked apparently. And Columbia was not even present. You might be prepared to scoff at the sport, but I admit that I am not (or will ever be) capable of what they are. Leaping, diving, throwing... and all with an astounding amount of speed and precision. Yea, I really should go for a run today.
Research-wise, I made the welcome discovery that despite the amount of effort I put into it, I should not have spent my last week translating the MCDI into Hindi. Lesson learned. In any case, I'm reading more papers on the actual implementation of a certain model. They allocate a 6x11 matrix for every syllable, which is troubling and something I'm not sure I want to replicate. We're still working out the kinks of Emergent and its somewhat poorly documented interface. I also got to run my first kid for the behavioral project. Just glad that it was my first and just a trial run. Only a few weeks left! A bit worried, but things will speed up.
Back to Top
It was really beautiful in Boulder this week, living up to its famed meteorological perfection. On another positive note, I was able to complete several of the layers for the neural network. So, work is coming along, and my surroundings are gorgeous. Good times.
Also, my friend, who's doing research at the Mountain Research Station at Niwot Ridge, came to visit me. I finally got a chance to walk over to Chautauqua and enjoy the alpine bliss that is only ten blocks from my dorm. Really unbelievable that students get to go to school here. It's completely stunning. Plus, I got to touch all the wares at the Jewish Festival. Knish, it was good. Altogether a good, productive week. Hopefully by the end of the next I'll have a model up and running, bug-free.
Back to Top
After scuffles with Python and Excel, I finally got a model that ran without errors. However, there is still "parameter tweaking" left to do before experiments can begin. I also managed to run a few more kids; it's getting easier and we've shortened the trials so that the children (and we ourselves) do not get bored out of their minds. Though I do think we had more sophisticated concerns when we modified that bit. Things are chugging along.
Back to Top
I tried sorting through all the LaTeX editors (LyX, LaTeXit, TeXShop, etc.) I've accumulated over time, presumably while I was trying to procrastinate on Discrete Math. I have no idea really. Anyway, currently quite pleased with emacs + Skim Auto update. Simple and no buttons. I should be finished with the introduction and methods, basically the scaffolding, of my report by the end of the weekend.
I ran several more kids, and I am still in the process of working out the kinks in the model, which seems to stall after several (hundred) epochs of steadily decreasing error rates. It's not unlike having a child, and feeling inadequate and irrationally angry while they sheepishly fail at learning basic skills. Skills required to function. Ugh. This model, seriously.
Back to Top
More parameter tweaking. This is somewhat maddening. However, someone who is more familiar with Emergent was able to help us get the model to stop thrashing. It's also reassuring to know that other people find these networks hard to work with. Combined with the interface, it's a little too much, especially when my inclination is to find bugs. There really aren't any bugs to be found, just numbers to be played around with.
I've started writing up my final report, and I'm actually using the LaTeX environment I researched and set up (emacs + Skim). I recommend it, especially for LaTeX newbies like me.
Back to Top
It's the last week! The model, we've conceded, does not work as we want it to. However, it does run and will accept phonological input. It's setting phonology as the output that's the problem. I foresee more tweaking after my internship ends. The tweaking never ends, really, when it comes to neural networks.
I continued to write up my report, re-reading the literature I received in the beginning of my internship. I'm getting excited about the problem again, when confronted with what made it originally so interesting, rather than what makes it so continually frustrating (tweaking). Talked with Professor Colunga, and I'm content with the work I've completed, with an eye for work in the future. Report forthcoming.
I also went to Denver on Saturday night to see fireworks and the Colorado Symphony Orchestra. It was brilliant. I think Denver's just the right size; DC or NYC would have overwhelming. All those people.
Back to Top
DREU asked me to document the five extra days that I worked to compensate for my tenth week. Although it's hard to remember many weeks after the fact, I believe on May 21st I completed a lot of reading, as I was in the literature stage of my project.
On June 4th, I was able to finish my MCDI. I remember quite distinctly staying up until 10pm on Saturday to finish the MCDI. I had promised Professor Colunga that I would finish by the end of the week, and my week that week technically went until Saturday 10pm.
June 18th mostly consisted of tweaking the parameters of the network, to little or no success. It's good to look at the learning rate, then the hidden layer parameters, then make sure that the matrices are sparse enough. It's a lot of heuristics, and it's very much like dealing with things in the real world, rather than the neat, (more) logical problems that CS usually presents.
On July 2nd and 3rd, I was anticipating the Fourth of July weekend, and ended up doing a lot of experiments with the model. We had decided that for DREU purposes, the model wouldn't be able to work completely properly, but that we would be able to approximate relationships (replace "reading" with "dictation" and replace "speaking" with "listening"). Mostly, there are still problems with the phonology layer.
Late July update: I am studying abroad in India at the moment, which somewhat accounts for the lack of updates. It's a little much to deal with, in addition to completing my DREU report: a new time zone, bureaucratic system, and culture. However, I bring this up to give an excuse for my tardiness, as well as to say that I will be continuing my research with Professor Colunga here in India, under the supervision of a professor here in Hyderabad. Because ten weeks is such a short time to learn a new field and produce publishable work in it, I'm grateful for the opportunity. I hope to update this with new developments.
Back to Top