Explainable AI
Explainable AI
Embedded Files
I have broad interests on explaining and interpreting deep learning models which is essentially considered as a black-box.
Specifically, I have been working on extracting human interpretable policies from various control environments as the form of mathematical equations or decision trees. Currently, I am trying to understand how to understand pre-trained RL agents operates using perception based XAI techniques such as grad-CAM( Gradient-weighted Class Activation Mapping) \, CRP(Concept Relevance Propagation) and TCAV(Quantitative Testing with Concept Activation Vectors). I am also trying to interpret Graph-based drug prediction model using more general methods such as LRP (Layer-wise Relevance Propagation). By understanding AI model, I ultimately have an (ambitious) goal on making the model as interpretable, tractable, transparent, trustable, deployable and most importantly, generalizable.
I have broad interests on explaining and interpreting deep learning models which is essentially considered as a black-box.
Specifically, I have been working on extracting human interpretable policies from various control environments as the form of mathematical equations or decision trees. Currently, I am trying to understand how to understand pre-trained RL agents operates using perception based XAI techniques such as grad-CAM( Gradient-weighted Class Activation Mapping) \, CRP(Concept Relevance Propagation) and TCAV(Quantitative Testing with Concept Activation Vectors). I am also trying to interpret Graph-based drug prediction model using more general methods such as LRP (Layer-wise Relevance Propagation). By understanding AI model, I ultimately have an (ambitious) goal on making the model as interpretable, tractable, transparent, trustable, deployable and most importantly, generalizable.
Optimizing discrete symbolic equation using Reinforcement Learning
1. Developing a framework that leverages reinforcement learning for symbolic regression.
2. Specifically, we use an RNN to emit a distribution over tractable mathematical expressions and employ a risk-seeking policy gradient to train the network to generate better-fitting expressions.
3. Applying the developed framework for interpretability, focusing on control dynamics and medical domain.
Optimizing discrete symbolic equation using Reinforcement Learning
1. Developing a framework that leverages reinforcement learning for symbolic regression.
2. Specifically, we use an RNN to emit a distribution over tractable mathematical expressions and employ a risk-seeking policy gradient to train the network to generate better-fitting expressions.
3. Applying the developed framework for interpretability, focusing on control dynamics and medical domain.
Selected Publications (* Equal contribution)
Sookyung Kim*, Mikel Landajuela Larma*, Brenden K. Petersen*, "Deep Symbolic Policy: Learning Interpretable Symbolic Policy with Reinforcement Learning". Proceedings of the International Conference on Machine Learning (ICML), 2021. [Oral presentation]
Brenden K. Petersen, Mikel Landajuela Larma, Claudio Santiago, Sookyung Kim. "Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients". Proceedings of the International Conference on Learning Representations (ICLR), 2021.[Oral presentation]
Sookyung Kim*, Joanne T. Kim*,Brenden K. Petersen*. "An interactive visualization platform for deep symbolic regression." Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence. 2021.
Explainable Drug DiscoverY (EDDY)
1. Develop an end-to-end explanation framework for graph neural network (GNNs) based deep learning models used in various drug discovery tasks.
2. Adopt developed framework for the research in three crucial sub-research areas of XAI:
(i) extracting model's interpretations using LRP (Layer-wised Relevance Propagation
(ii) evaluating multiple interpretations, and
(iii) developing collaborative visual tools to engage domain scientists in the process of adjudicating explanation quality.
Explainable Drug DiscoverY (EDDY)
1. Develop an end-to-end explanation framework for graph neural network (GNNs) based deep learning models used in various drug discovery tasks.
2. Adopt developed framework for the research in three crucial sub-research areas of XAI:
(i) extracting model's interpretations using LRP (Layer-wised Relevance Propagation
(ii) evaluating multiple interpretations, and
(iii) developing collaborative visual tools to engage domain scientists in the process of adjudicating explanation quality.
Selected Publications
Sookyung Kim, Subhashis Hazarika, Hiroshi Kajino, Md Mahim Anjum Haque. "Analyzing Chemical Structure of Drug using Molecular Hypergraph Grammar" Submitted to the Machine Learning for Structural Biology Workshop at NeurIPS, 2023.
Mahim Anjum Haque, Subhashis Hazarika and Sookyung Kim. Visual Framework to Evaluate Explanation Methods for Graph Neural Network based Drug Screening Models. Submitted to the demonstration track on AAAI, 2023.
Patents
Sookyung Kim, Subhashis Hazarika "Analyzing Chemical Structure of Drug using Molecular Hypergraph Grammar" US-Patent Pending
Subhashis Hazarika and Sookyung Kim. Visual Framework to Evaluate Explanation Methods for Graph Neural Network based Drug Screening Models.US-Patent Pending
Page updated
Google Sites
Report abuse