Joon Kim

I am a machine learning research scientist at DynamoFL, working on developing compliant AI solutions for enterprises.

I completed my PhD at the Machine Learning Department of Carnegie Mellon University, where I was advised by Ameet Talwalkar. My research focused on methods that can assist our understanding of complex machine learning models, and their application to useful downstream tasks involving human users.

Prior to my graduate studies, I received a Bachelors of Science degree in Computer Science at Caltech.

[CV] [GoogleScholar]

Publications

	Assisting Human Decisions in Document Matching Joon Sik Kim, Valerie Chen, Danish Pruthi, Nihar B. Shah, Ameet Talwalkar Transactions on Machine Learning Research (TMLR), 2023. proceedings / code / demo We conduct a crowdsourced study to test different methods that can be useful for human decision makers in document matching, a practial application including academic peer review.
	Bayesian Persuasion for Algorithmic Recourse Keegan Harris, Valerie Chen, Joon Sik Kim, Ameet Talwalkar, Hoda Heidari, Zhiwei Steven Wu Neural Information Processing Systems (NeurIPS), 2022. proceedings We study the problem of offering algorithmic recourse without requiring full transparency about the model and how a decision maker can incentivize mutually beneficial action to the decision subjects.
	Sanity Simulations for Saliency Methods Joon Sik Kim, Gregory Plumb, Ameet Talwalkar International Conference on Machine Learning (ICML), 2022. (Spotlight) XAI4CV Workshop, CVPR 2022. proceedings / slides / poster / code We propose a benchmark study on the limitations of leading saliency methods via several stylized tests of identifying correct sets of features based on different model reasoning.
	Interpretable Machine Learning: Moving From Mythos to Diagnostics Valerie Chen, Jeffrey Li, Joon Sik Kim, Gregory Plumb, Ameet Talwalkar Communications of the ACM. Vol. 65 Issue. 8, August 2022. Workshop on Human in the Loop Learning (HILL), ICML 2021. proceedings We survey the current landscape of a field of interpretable machine learning and ways to concretely address existing issues and disconnects.
	Efficient Topological Layer based on Persistence Landscape Kwangho Kim, Jisu Kim, Manzil Zaheer, Joon Sik Kim, Frederic Chazal, Larry Wasserman Neural Information Processing Systems (NeurIPS), 2020. proceedings / code We introduce a novel topological layer for deep neural networks based on persistent landscapes, which is able to efficiently capture underlying toplogical feautres of the input data, with stronger stability guarantees.
	FACT: A Diagnostic for Group Fairness Trade-offs Joon Sik Kim, Jiahao Chen, Ameet Talwalkar International Conference on Machine Learning (ICML), 2020. proceedings / blog post / slides / code We proposed a general framework of understanding and diagnosing different types of trade-offs in group fairness, deriving new incompatibility conditions and post-processing method for fair classification.
	Automated Dependence Plots David Inouye, Leqi Liu, Joon Sik Kim, Bryon Aragam, Pradeep Ravikumar Uncertainty in Artificial Intelligence (UAI), 2020. Safety and Robustness in Decision Making Workshop, NeurIPS 2019. proceedings / code We introduce a framework for automating a search for a partial dependence plot which best captures certain types of model behaviors, encoded with customizable utility functions.
	Representer Point Selection for Explaining Deep Neural Networks Chih-kuan Yeh, Joon Sik Kim*, Ian E.H. Yen, Pradeep Ravikumar Neural Information Processing Systems (NeurIPS)*, 2018 proceedings / blog post / poster / code We propose a method that decomposes a deep neural network prediction into a linear combination of activation values of training points, in which the weights (called representer values) allow intuitive interpreation of the prediction.
	A Rotation Invariant Latent Factor Model for Moveme Discovery from Static Poses Matteo Ruggero Ronchi, Joon Sik Kim, Yisong Yue IEEE International Conference on Data Mining (ICDM), 2016 proceedings / project page / slides / poster / code / dataset We propose a method to discover a set of rotation-invariant 3-D basis poses that can characterize the manifold of primitive human motions, from a training set of 2-D projected poses obtained from still images taken at various camera angles.