I am a Ph.D. student in Computer Science at Columbia University, advised by Elias Bareinboim, working on generalizability in machine learning from a causal perspective.
I use the language of causal inference to formalize the inductive biases that enable the agents to make cross-population inferences. Through this lens, I study the theoretical limitations of learning in tasks such as domain generalization, domain adaptation, and transfer learning.
Before joining Columbia, I completed my undergraduate degree in Computer Science and Economics at Sharif University of Technology, Tehran, Iran.
Email: kasra at cs dot columbia dot edu
If you'd like to work with me as a mentee, please check out this note.
In submission
Does causal structural similarity between source and target domains explain few-shot learnability? We develop a structure-agnostic procedure that attains fast adaptation rates (i.e., truly few-shot learning) in situations where zero-shot generalization is possible via a transportability based structure-informed algorithm that leverages causal assumptions.
We show that neural causal models can be employed to compute tight upper-bounds for the error of classifiers in the unseen domains using causal diagrams as expert knowledge. We also introduce Causal Robust Optimization (CRO), an iterative scheme for finding the best worst-case classifier within a hypothesis class.
We introduce transportable representations, a causal framework for analyzing domain generalization. Our algorithm determines whether label predictions given a representation can be computed from source domains using structural assumptions or data alone. We prove a graphical-invariance duality and show that causal and non-causal features can both support generalization, unifying prior approaches under transportability theory.
We tackle causal discovery across multiple domains with varying noise but shared causal functions. Assuming absence of confounders, we show that bijective generation mechanisms enable recovery of the parent set of a target variable under weaker functional assumptions, unifying some of the existing identification results.
We employ an implementation of the MDL based on normalized maximum likelihood (NML) for granger-causal discovery in point processes. We introduce a Monte Carlo estimation technique to compute NML, and show that causal discovery in this setting is equivalent to an optimaol model selection scheme.