About Me
I am a Ph.D. student at the Edmond and Lily Safra Center for Brain Sciences (ELSC) at The Hebrew University of Jerusalem. My research focuses on scales of description in science and physics, exploring how different levels of abstraction relate to one another. Using machine learning tools, particularly artificial neural networks (ANNs), I study how high-dimensional information can be transformed into meaningful lower-dimensional representations.
My Research
I am interested in scales of description in science and physics. For example, we can view a person as a collection of atoms or as part of a community. How do these different scales relate? Advances in machine learning provide tools to study these questions.
Abstract Reasoning in Neural Networks
We tested an ANN's ability to extract a predictably changing feature from image sequences with various changing features. The ANN performed well on intelligence tests requiring this ability, validating its use in modeling the extraction of abstract theories.
Theory Adaptation in Response to Contradictions
We studied how theories change in response to observations that contradict them. Our ANN's choice of resolution depended on the inconsistency's magnitude, highlighting flexibility in forming descriptions.
Analogies Between Scales of Description
We are studying how ANNs form analogies in general, particularly between scales of description. This work relates to the Renormalization Group framework in physics, which preserves theory structure across scales.
Publications
Two pathways to resolve relational inconsistencies
We studied how theories change in response to observations that contradict them. Our ANN's choice of resolution depended on the inconsistency's magnitude, highlighting flexibility in forming descriptions.
Untrained neural networks can demonstrate memorization-independent abstract reasoning
We tested an ANN's ability to extract a predictably changing feature from image sequences with various changing features. The ANN performed well on intelligence tests requiring this ability, validating its use in modeling the extraction of abstract theories.