People are incredibly flexible and efficient inductive reasoners. How can we model this ability computationally? In my thesis, I propose two hypotheses: people may operate over an incredibly vast language which is made tractable via a strong bottom-up proposal model, or people may learn task-specific languages to reason over.
Proposes a rational process model for how people do program induction and shows that it better fits human performance than existing models with significantly less samples.
Language models show incredible cross-domain program induction ability but lack robustness and planning abilities. We show that by using language models to learn a domain-specific language, we can redress these deficiencies, improving program induction performance beyond both language model performance and program induction performance using a manually-designed DSL.
Compositionality, the capacity to understand and generate complex ideas by combining simpler ones, is a critical attribute of human intelligence and plays a pivotal role in sophisticated problem-solving tasks. In this thesis proposal, I outline projects that redress the deficiencies of several limiting assumptions of prior work.
Combines the vector-quantization of Vector Quantized-Variational Autoencoders (VQ-VAEs) with the hyperbolic embedding space of hyperbolic Variational Autoencoders (hVAEs) to train a Hyperbolic Vector-quantized Variational Autoencoder in a step towards learning discrete hierarchical representation.
Many accounts of how people form meaning rely on people's ability to use compositionality, the ability (roughly) to combine simple concepts into more complex ones. Prior work has shown that 4-year-olds can use compositionality in visual tasks, but failed to show the same ability in younger infants. In this project, I wrote a mock paper about a study I designed to test whether young infants can use compositionality in a visual task.
It is well known that humans have incredibly flexible and general object recognition capabilities, especially in comparison to machine vision systems. In an attempt to learn more about what might be causing this disparity in performance and to improve machine vision systems, we explore whether constraining a machine vision system to be more human-like improves the performance of the system in learning new tasks.
LARC (Language-annotated Abstraction and Reasoning Corpus) is a dataset I collected with Yewen Pu and other post-doctorates in the MIT Computational Cognitive Sciences lab. ARC is a benchmark for abstract thinking ability in programs, and our dataset augments ARC with hundreds of natural-language descriptions which capture the entirety of the task, collected through a multi-person communication game.
Art created using code. Varies from combining GANs and diffusion models with CLIP to generating simple flow fields. Since the beginning of 2022, I have tried to create one piece a day.
For my final project for the class Advanced Natural Language Processing (6.864), I worked with Lowell Hensgen to explore how individual neurons contribute to the final answer of large multimodal models, as well as how model explainability relates to model performance on a suite of tasks.
For my final project for the class Computational Cognitive Science (9.66), I explored computationally inferring the moral values of different cultures using a hierarchical Bayesian model. To do this, I focused on The Moral Machine domain, a database of tens of thousands of decisions made about moral dilemmas from individuals around the world.
API and app that enables anyone to create and customize stock-trading algorithms without code, just by tuning parameters. The program runs a plethora of trading algorithms on the cloud that are customizable by the end-user without needing to write any code.
Li–O2 batteries can provide significantly higher gravimetric energy density than Li-ion batteries, but their practical use is limited by a number of fundamental issues associated with oxidizing discharge products. In this work, we looked at using different solvent mixtures to tune the stability of inorganic redox mediators in Li-O2 batteries.