Ben Prystawski

Ben Prystawski

Computational Cognitive Science Student

University of Toronto

About Me

I am a fourth-year undergraduate student in computer science and cognitive science student at the University of Toronto, conducting research in computational cognitive science and computational linguistics.

I work in the labs of Yang Xu, Joseph Jay Williams, Daphna Buchsbaum, and Falk Lieder and I am applying to graduate school in fall 2020, hoping to start a PhD in Fall 2021.

Interests

  • Cognitive Science
  • Computational Linguistics
  • Bayesian Statistics
  • Probabilistic Programming
  • Reinforcement Learning

Education

  • B.Sc. in Computer Science and Cognitive Science (expected), 2021

    University of Toronto

Projects I am Involved in

Gendered Language in Childhood

How do parents speak differently to boys compared to girls? And do gender differences in children’s speech reflect the differences in speech from parents? I am investigating these gender differences in Professor Yang Xu’s lab using the CHILDES corpus and word embeddings like Word2Vec, GloVe, and fastText to measure gender associations prevelant in broader society.

Explore-Exploit Tradeoff in Causal Inference

When inferring causal relationships, we must balance exploring hypothesis space by trying completely new hypotheses with exploiting the effectiveness of our current hypothesis at explaining the data by making minor revisions. I am working in Professor Daphna Buchsbaum’s lab on modelling the process of hypothesis revision in causal inference using particle filters and Markov Chain Monte Carlo over a structured space of rules. We are also gathering experimental data from both children and adults to investigate how the tendency to explore or exploit in causal inference changes over development.

Hypothesis Testing for Adaptive Experiments

It is often valuable to be able to run experiments adaptively, particularly in settings such as educational experiments and clinical trials. That is, rather than simply assigning participants to conditions uniformly randomly, assigning them based on the existing evidence that the condition is most helpful. This is most often done with bandit algorithms. However, the data gathered using bandit algorithms often leads to higher false-positive rates and reduced statistical power when standard hypothesis tests are used. In this project, supervised by Professor Joseph Jay Williams, I work on modifying bandit algorithms and statistical hypothesis testing to enable experimenters to assign more people to better experimental conditions at minimal cost to statistical rigour.

Modelling Human Goal Pursuit

Setting and pursuing goals are fundamental aspects of human cognition. By definition everyone wants to achieve their goals, but people often fall short of the goals they set for themselves. This project, supervised by Dr. Falk Lieder, used resource-rational analysis to understand how cognitive constraints shape human goal-directed behaviour. We draw upon ideas from cognitive psychology, behavioural economics, and reinforcement learning to better understand the role of limited cognitive resources in human goal pursuit.