Have a UROP opening you would like to submit?
Please fill out the form.
Developing neurosymbolic approaches to a visual human reasoning task
9: Brain and Cognitive Sciences
Dr. Andrzej Banburski: email@example.com
Despite huge advances made in the application of deep neural networks to a wide variety of tasks, neural network approaches suffer from a thirst for data and an inability to generalize to new tasks. The recently introduced ARC (Abstraction and Reasoning Corpus) dataset is a new benchmark that seeks to measure human-level intelligence in machine learning models, focusing on types of reasoning humans excel at which machines find difficult. We are developing a system to tackle the ARC dataset with a combination of neural networks and program synthesis techniques. Your role: You will help us design, implement, and experiment with novel approaches to the ARC dataset. Things to do include: designing a systematic approach to solving ARC, implementing basic operations needed to solve, running experiments, and contributing ideas to the direction of the project. You will be collaborating with a group of 4-5 other researchers also working on the project. Estimated hours per week: 6+ Time period: Fall, with extension into IAP or spring welcome. Remote? Yes
Your background: - Interested in the intersection of neural networks and symbolic reasoning - Coding in Python - Eager to contribute - Any experience with program synthesis, functional programming, deep learning, or neuroscience is a plus.