Applied research
Long-term programs connect SAIL's machine learning research to public, scientific, clinical, and industrial needs.
14 active programs
Current program · 01
The XAI center is developing new or modified machine learning techniques to produce explainable AI models. The explainable AI models provide effective explanation for human to understand the reasons of decision making via machine learning and statistical inference based on real-world data. The center highly contributes to Medical and Financial industry, which obtains high advantage by adopting AI but high risk follows without explainability.
Current program · 02
This project focuses on the development of a specialized AI agent designed to support professional weather forecasters in their operational duties. The primary goal is to enhance the decision-making process for forecasters, thereby reducing their workload and helping to minimize the impact of natural disasters.
Current program · 03
Current program · 04
Interpretable models for acute kidney injury risk prediction and clinical prevention workflows.
Current program · 05
Models that incorporate changing policies and financial regulations, including the Basel Accords, into explainable decisions.
Current program · 06
Financial-related policies and regulations are amended and established according to changing times or economic conditions. In this project, we develop explainable AI models that make decisions by reflecting these changing policies and financial regulations (e.g., Basel Accords).
Current program · 07
Methods for understanding relationships inside Transformer modules and making robust inferences on unseen data without further training.
Current program · 08
This project aims to address the limitations of neural scaling laws by developing methods for robust inference using the knowledge captured by Transformer models. The first phase focuses on understanding and explaining the relationships between the learned information within Transformer modules. In the second phase, the goal is to create algorithms that enable accurate inference on unseen data without further training, by effectively integrating existing knowledge with new information.
Current program · 09
Robust, controllable, and unbiased trustworthy AI with stable behavior across diverse environments.
Current program · 10
Current program · 11
Current program · 12
Current program · 13
Current program · 14
User-centered explanations for deep weather forecasting systems.
Planning and learning for coordinated autonomous systems.
A model-agnostic method for post-hoc prosody control and pronunciation correction without retraining.
A game-theoretic attribution rule that accounts for feature interactions in complex black-box models.
Handling large streaming data is essential for various applications such as network traffic analysis, social networks, energy cost trends, and environment modeling. However, in general, it is intractable to store, compute, search and retrieve large streaming data. This research solve the issues by the “Locally Exchangeable Blocks” in which sample data points are exchangeable each other. Thus, we can dynamically adjust the sampling rates by maintaining reliable statistical inference.
Kalman Filtering (KE) is a computational tool with widespread applications in robotics, financial and weather forecasting, environmental engineering and defense. The Kalman Filter (KF) recursively estimates the state variables of a dynamic system with cubic time matrix multiplication and inversion operations. Here, we proposed Relational Gaussian Models to represent and model dynamic systems with large numbers of variables efficiently, and devised an exact lifted Kalman Filtering algorithm which takes only linear time in the number of random variables at every timestep. We prove that our algorithm takes linear time in the number of state variables even when individual observations apply to each variable.
An efficient video retrieval system is essential to find relevant video contents from a large set of video clips, which typically contain several heterogeneous video clips to match with. In this project, we introduce a content-based video matching system that finds the most relevant video segments from video database for a given query video clip. For the video retrieval task, we propose a novel video matching called Spatio-Temporal Pyramid Matching (STPM). Considering features of objects in 2D space and time, STPM recursively divides a video clip into a 3D spatio-temporal space like a pyramid and compares the features in different resolutions.
Data from sensors in a nuclear power plant are sampled periodically and trends are compared against the steady-state condition to determine if an anomaly exists. If an anomaly is detected, the proposed diagnosis system (PRODIAG) attempts to identify the cause through the qualitative reasoning process that relates faults to sensor trends and knowledge of how plant components are connected. The algorithms in the system only needs to know the Piping and Instrument Diagram for the plant making the approach highly portable across differing nuclear plants. This work is a collaborative effort with the Nuclear Division of Argonne National Laboratory.