Ever since reading Hitchhiker's Guide and watching Monty Python and the Holy Grail as a kid, I've been a huge fan of British comedies. I wanted to see if I could use an LSTM recurrent neural network to figure out the differences between British and American humor. To teach a network how to distinguish the two, I trained it to tell apart episodes of the US and UK Office and then tested whether it could predict whether other shows like Parks and Rec, Extras, Friends and Fawlty Towers were American or British. The Git.

Turkish Delight: Sentiment classification and topic modeling of complaints on the mturk subreddit (+mturk tips)

I use Mechanical Turk for my experiments a lot (see here, here, here, here). And over the years, I've picked up a good amount of experience on how to create tasks for mturk. As psychologists who never have any contact with their subjects, it can be easy to ignore their comments and complaints (if you want a nice shock, see this). But mturkers have an active community and they regularly talk with each other about problems they've had with HITs and requesters.

To better understand mturkers concerns, I scraped responses from the mturk subreddit, used a Naive Bayes sentiment classifier to identify people's complaints and then organized them into topics using LDA. Based on the topics I got, I wrote tips on how to create better HITs on Mechanical Turk. The Git.

Using topic modeling and NLP to improve multiple choice question design

Although multiple choice questions on exams are meant to probe students' knowledge, poorly worded questions can require students to solve complex logic puzzles. How can teachers design multiple choice questions that focus on assessing students' learning? I use Latent Dirichlet allocation topic modeling and NLP analyses to measure how multiple choice questions are worded and then use mixed effects modeling to determine how wording affects people's performance on multiple choice questions. As a plus, we can use mixed effect models to predict the difficulty of new questions' wording. The Git.

Eatrovert: Discovering unique cuisines in new cities

It can be difficult to find cool, new restaurants when you're visiting a new city. I scraped data from Yelp and compared the distribution of restaurants between different cities to help foodies discover new cuisines when they're travelers. Above, consistent with common sense and my own (delicious) personal experience, Eatrovert recommends a New Yorker should go to San Diego to try Mexican food in Old Town. Enjoy! The Git.

Cosmopolitan: Measuring people's awareness of the world

As the world becomes more interconnected, the risk of cultural misunderstandings increases. Understanding what countries people are interested in and know about can consequently help us determine where there might be a high risk of faux pas. We measured people's knowledge of other countries by using Google Trends to compare how much countries are searched for and search for each other. In the figure above, we examined how much countries ("searcher") searched other countries ("searchee") (darker blue indicates the searcher Googled the searchee more). For example, the dark blue of the top-row indicates people search for the US frequently and the dark blue English column demonstrates that the English are pretty cosmopolitan and seek out news about the rest of the world. The full Git.

Who matters?: Revealing biases in interpretations of disasters in the news

Full interactive map

In the wake of the November 2015 Paris attacks that killed 128 people, Facebook users across the world expressed their support for Parisians by adding French flag filters to their profiles and posting statuses like "Pray for Paris". Yet some observers noted that at the same time, more than 2000 Nigerians were killed by the Boko Haramand an al-Shabab attack killed 147 Kenyans but received little media attention. These imbalances in coverage raise the question, what does it take for the general public to care about a tragedy? Do people expect more deaths in countries outside the general "Western world" such that they are not shocked unless there is a sufficiently terrible catastrophe? Here, we quantify how large a catastrophe must be to get Americans' attention. Darker red indicates that people expect more casualties in the country, such that people expect a lot of casualties in Nepal, Haiti and Rwanda. The full Git.

Simplifying explanations of income inequality

Income inequality has become a critical issue in American politics. Yet, it can be hard to conceptualize how much income inequality there is. If I ask you how much someone in the top 1% makes, your guess is probably going to be pretty far off. Consequently, the ability to frame inequality in concrete, comprehendible ways can help policymakers construct more convincing arguments for reform. people a make arguments for policy reform. Here, I wanted to see if I could get a more intuitive measure of income inequality, one that reflects the kinds of social judgments we make everyday: If I meet someone of Race X and Gender A, what is the probability that my income is higher? For example, the biggest inequality in the graph above demonstrates that a random Caucasian male has a probability of ~75% of making more than a Hispanic female. The full Git.

Visual memory iterated learning videos

Videos of chains from "Structured priors in visual working memory revealed by iterated learning", Cognitive Science Society (2015).

Consider a game of Telephone, where one child thinks of a word, whispers it to the next child, and then the next child has to pass on the word they heard. However, they have uncertainty about the meaning of the word. To compensate for their uncertainty, they rely on their prior expectations about what sort of words the other child might've spoken. As a result, as the word is passed between more and more children, it will increasingly resemble the children's priors.

Here we examined what priors people have about the spatial arrangement of objects. Each trial, the participant briefly saw and then recalled the locations of 15 circles. The next participant saw and then recalled the previous participant's responses. Over time, participants converged towards complex, sophisticated spatial patterns.

Iterated learning task

In this demo, you and some collaborators (AKA friends) can generate an iterated learning chain, just like in "Structured priors in visual working memory revealed by iterated learning", Cognitive Science Society (2015).

Super Smith Bros: Kevin's Dissertation Dash!

To commemorate Kevin Smith leaving the Vul Lab for his post doc, I used my vismem toolbox to make a simulation (i.e., awesome, totally accurate video game) of Kevin's time here at UCSD. If I ever get around to it, the expansion pack will include some wild speculations about Kevin's future, most likely involving throwing fireballs and jumping on turtles.

vismem Toolbox

This js file contains a suite of functions I wrote for creating basic visual memory experiments using HTML's canvas.