Projects

This page presents the projects I have been involved in to date. Since around 2003, I have participated in numerous AI (machine learning and data mining) projects at IBM Research spanning various industries. Our activities demonstrate a unique synergy between fundamental research and practical business applications.

Current project

Currently, I am serving as the Head of Data Science at IBM Semiconductors in IBM’s Research Division, located at the IBM T. J. Watson Research Center in New York, USA. My main focus is enhancing the efficiency of semiconductor manufacturing, particularly IBM’s 2 nm fabrication process, through the utilization of machine learning (ML) technologies. I am responsible for leading strategic research initiatives including WIP simulation, advanced process control, process scheduling, and fault detection.

Previous projects

Causal discovery in complex domains (2022-2023)

In this project, my goal was to reconstruct the theory of causal analysis from an explainable AI (XAI) perspective. There are approximately four major approaches to causal analysis, which depend on the definition of causality. Sometimes, these definitions can be more philosophical than mathematical. The approaches include:

  1. Inference on a Bayesian network
  2. Structural equation analysis
  3. Intervention analysis
  4. Granger causal analysis

While each approach has a strong theoretical foundation, it is worth noting that there is a gap between the problem setting and the business requirements. In practical terms, business end-users typically seek the ability to prioritize tasks for more efficient and effective action. This aspect has not been extensively discussed in the machine learning community’s literature on causal analysis.

On the other hand, XAI research has encountered its own challenges in terms of practical applicability. Early XAI research mainly focused on the psychological barriers that humans face when accepting machine-generated answers. Even to this day, the majority of studies, including those conducted by the IBM Trusted AI group, implicitly assume deep image classification as the primary task, while giving limited consideration to the real business requirements for actionable insights.

To complement the research topics addressed by the Trusted AI group, I primarily led three projects:

  1. Event causal learning using deep Hawkes models, in collaboration with Mr. Dongxia Wu and his advisor Prof. Rose Yu from UC San Diego.
  2. Transformer-based anomaly diagnosis of industrial systems, in collaboration with Dr. Pin-Yu Chen, one of the most prolific scientists of IBM Research, Mr. Jokin Labaien and his advisor Dr. Ekhi Zugasti and Dr. Xabier De Carlos from Ikerlan/Mondragon University.
    • See, e.g., J. Labaien, T. Idé, P.-Y. Chen, E. Zugasti, X. De Carlos, “Diagnostic Spatio-temporal Transformer with Faithful Encoding,” Knowledge-Based Systems, accepted, 2023.
  3. Anomaly attribution theory, in collaboration with Dr. Naoki Abe of IBM Research.
    • See, e.g., T. Idé and N. Abe, “Generative Perturbation Analysis for Probabilistic Black-Box Anomaly Attribution,” KDD 2023.

I was also involved in a project on graph neural networks, led by Dr. Georgios Kollias, one of my talented colleagues, which resulted in ICASSP 23 and AAAI 22 papers.

AIOps (2020-2021)

This is another project that falls into the category of condition-based maintenance of industrial systems. In this project, we dealt with a new type of data: stochastic events from computer systems (computer alerts, warnings, etc.). I developed a new causal analysis framework called the Granger-Hawkes model to address the question of “who caused this?” for event sequences (see the figure below). We applied the model to event grouping for alert and working events from IBM’s cloud system.

The details of the technology and use-case scenarios are described in my NeurIPS paper. The technology has been incorporated into the IBM Watson Core library.

Tsuyoshi Idé, Georgios Kollias, Dzung T. Phan, Naoki Abe, “Cardinality-Regularized Hawkes-Granger Model,” Advances in Neural Information Processing Systems 34 (NeurIPS 21, Dec 6-14, 2021, virtual), pp.2682-2694 [slides, poster].

Explainable AI for IoT (2018-2020)

We collaborated with the IBM Watson IoT business unit. In Research, I led the project of developing methods and algorithms that explain unusual events detected by a black-box prediction model y=f(x), with the main application being building energy management. I invented a fundamentally new approach to explain detected anomalies, which has been implemented in two IBM products, IBM Tririga Building Insights and IBM OpenScale, along with an uncertainty quantification method.

My AAAI paper provides a concise summary of the technology we developed:

Tsuyoshi Idé, Amit Dhurandhar, Jiri Navratil, Moninder Singh, Naoki Abe, “Anomaly Attribution with Likelihood Compensation,” In Proceedings of the Thirty-Fifth AAAI Conference on Artificial Intelligence (AAAI 21, February 2-9, 2021, virtual), pp.4131-4138 [slides, poster, official version].

AI for Blockchain (2018)

In 2015, IBM published a white paper beautifully titled “Device Democracy” that highlighted the significant impact of the concept of Blockchain in the IT industry. I proposed new research and business directions to enhance Blockchain from a transaction management system to a decentralized platform for value co-creation among participants, with machine learning playing a critical role.

business directions to advance Blockchain from a mere transaction management system to a decentralized platform for value co-creation among participants, in which machine learning plays a critical role.

Although my business proposal did not succeed internally, it received significant attention in the research community. I presented an invited talk at the 2021 IEEE International Symposium on Blockchain titled “Decentralized Collaborative Learning with Probabilistic Data Protection.” The extended abstract of my talk was published as a full conference paper in the proceedings of the prestigious IEEE International Conference on Smart Data Services (SMDS 21).

Tsuyoshi Idé, Rudy Raymond, “Decentralized Collaborative Learning with Probabilistic Data Protection,” In Proceedings of the 2021 IEEE International Conference on Smart Data Services (SMDS 21, September 5-10, 2021, virtual), pp.234-243 [slides, IEEE Xplore].

The technical concept was also featured in a paper presented at the 28th International Joint Conference on Artificial Intelligence (IJCAI 19), one of the top AI conferences in the world.

Tsuyoshi Idé, Rudy Raymond, Dzung T. Phan, “Efficient Protocol for Collaborative Dictionary Learning in Decentralized Networks,” Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI 19, August 10-16, Macao, China), pp.2585-2591 [slides, poster].

Smarter Manufacturing (2015-2017)

This project was a continuation of the sensor data analytics project I initiated in Japan a decade ago. Since then, the world has begun to recognize the significant potential of AI in the manufacturing industry. In the wake of this realization, numerous “new” business proposals have emerged, using trendy terms like Industry 4.0 and Smarter Manufacturing came out. Leveraging the advanced machine learning algorithms we developed earlier, I outlined the fundamental architecture for a solution suite in this evolving domain. 

For a general overview of my work, including my projects in the US, please refer to the presentation slides of my invited talk at an international conference:

Tsuyoshi Idé, “Recent advances in machine learning from industrial sensor data,” The 12th ICME International Conference on Complex Medical Engineering (CME 2018, September 6-8, 2018), Shimane, Japan [slides].

Jayant Kalagnanam, Young Min Lee, Tsuyoshi Ide, Cross Industry Analytics Solution Library for Resource and Operations Management, IBM Research Report RC25563 (WAT1510-012) October 6, 2015 [link].

Service Delivery & Risk Analytics (2013-2014)

In 2013, I was appointed as the manager of Service Delivery & Risk Analytics at IBM T. J. Watson Research Center in New York, USA. The primary objective of my team was to enhance IT service delivery using machine learning. Smarter IT service management was one of the three strategic focuses of Services Research in IBM Research during that time and was an area where machine learning could have a significant impact.

As the manager, I led two major initiatives. The first one involved the solution design phase of IT service delivery. I developed a new algorithm for predicting project risks based on data generated from the quality assurance process at IBM (refer to the figure). This work can be considered one of the earliest research contributions to the field of Explainability of AI.

The algorithm leverages a psychometrics approach called the item response theory and opened a new door to questionnaire data analytics. For details, see the KAIS paper:

Tsuyoshi Idé and Amit Dhurandhar, “Supervised Item Response Models for Informative Prediction,” Knowledge and Information Systems, p.1-23, 2016 [link, slides for related paper].

The other initiative was about the service delivery phase. In collaboration with the team members, I developed a text mining approach to IT service tickets. See, e.g.,

Kuan-Yu Chen, Ee-Ea Jan, Tsuyoshi Idé, “Probabilistic Text Analytics Framework for information Technology Service Desk Tickets,” Proceedings of the 14th IFIP/IEEE International Symposium on Integrated Network Management (IM 2015), 2015, pp.870-873.

Analytics & Optimization (2010-2013)

During 2010-2013, I served as the head of the AI department (the Analytics & Optimization group) at IBM Research – Tokyo. The first task I undertook was to establish a strategic research agenda that all team members were expected to contribute to:

  • Analysis of stochastic interacting systems
  • Analysis of industrial dynamic systems

For the analysis of stochastic interacting systems, our ultimate goal was to establish a methodology for analyzing complex systems such as societies, cities, and enterprises. I started several new and exciting projects across various industries. Among them, the Frugal Intelligent Transportation Systems for Kenya was one of the most successful projects and in fact gained a significant media coverage. The key concept was “Frugal Innovation.” Instead of relying on expensive social infrastructure in the traditional manner, we built a comprehensive ITS solely using inexpensive web cameras, which were enhanced with sophisticated image analysis and network inference algorithms (see figure below).

For more technical details, see:

Tsuyoshi Idé, Takayuki Katsuki, Tetsuro Morimura, and Robert Morris, “City-Wide Traffic Flow Estimation from Limited Number of Low Quality Cameras” IEEE Transactions on Intelligent Transportation Systems, 18 (2017) 950-959 [link, slides for related paper].

Since it is not possible to establish fully analytic models in complex systems, simulation technologies can be a powerful approach. However, one critical issue is how to validate the simulation results. To address this, I am interested in exploring how simulation could be combined with optimization technologies. For example, we may want to optimize the model of individual agents in a multi-agent traffic simulation using sophisticated machine learning technologies, possibly through a method similar to Bayesian optimization. It is my great honor to have an opportunity to launch a new Strategic Initiative in this area in the Math department of Global IBM Research.

For the analysis of industrial dynamic systems, major research topics include sensor data analytics and production optimization, which are particularly important in the Japanese market. My own research, including anomaly detection and trajectory analytics, plays a critical role in real production systems, such as ClassNK’s ship maintenance system.

Sensor Data Analytics (2005-2013)

After completing the Autonomic Computing project, I started a new project called Data Analytics for Quality Control, also known as Sensor Data Analytics. The goal of this project was to enhance the quality of products in the manufacturing industries by utilizing advanced analytics for sensor data. Initially, it was a one-person project, but it eventually grew into a major initiative across the entire corporation, thanks to the efforts of many colleagues.

One of the most crucial tasks during this period was the development of “proximity-based” anomaly detection algorithms. Specifically, in the SDM paper mentioned below, I introduced the sparse graphical model in the context of correlational anomaly detection.

Tsuyoshi Idé et al., “Proximity-Based Anomaly Detection using Sparse Structure Learning,” Proceedings of 2009 SIAM International Conference on Data Mining (SDM 09), pp.97-108 [slides].

The project was among the world’s earliest systematic efforts to leverage data-driven management in the Internet of Things (IoT) domain. When I launched the project, leveraging data in IoT primarily involved creating a centralized database to store attributes of various pieces of production equipment. Almost a decade after my initial proposal and the completion of successful customer projects, a broader audience has begun to recognize how combining advanced analytics with sophisticated database systems can revolutionize business practices.

Automated Analysis Initiative (AAI; 2003-2004)

This project was designed to create a comprehensive framework for sensor data analysis, with a specific focus on the automotive industry. Unlike previous initiatives, this project had a well-defined research agenda. I introduced the innovative concept of change-point correlation, a critical addition to our framework that later became integral to the IBM Parametric Analysis Center (see the SDM paper below for details). This method is particularly effective in managing the heterogeneity common across various sensor data types, such as temperature and pressure, found in industrial physical systems. The success of this project laid the groundwork for my subsequent endeavor, the Sensor Data Analytics project.

Tsuyoshi Idé, “Knowledge Discovery from Heterogeneous Dynamic Systems using Change-Point Correlations,” Proceedings of 2005 SIAM International Conference on Data Mining (SDM 2005), pp.571-576 [slides].

Autonomic Computing (2002-2003)

This project represented a company-wide effort to address the increasing complexities within computer systems. Transitioning from a vastly different field of physics/optics, I was new to this area. Nevertheless, I succeeded in establishing a research agenda that proved to be highly relevant for the field: anomaly detection for system monitoring. The resulting KDD paper, my inaugural publication in computer science, emerged from this work.

The paper is well-known to be one of the first works of subspace-based anomaly detection and enjoys 300+ citations as of 2022. The paper is also one of the first works that leverage directional statistics for scoring anomalies.

Tsuyoshi Idé and Hisashi Kashima, “Eigenspace-based Anomaly Detection in Computer Systems,” Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD 2004), pp. 440-449 [slides].

Collimated backlight (2000-2001)

The objective of this project was to create a backlighting system capable of achieving the highest light utilization in the world. Under the leadership of Dr. Yoichi Taira, our team developed an innovative backlighting technique known as “collimated backlight.” However, the ultra-efficient design presented a significant challenge: maintaining luminance uniformity across the display’s entire surface proved to be exceedingly difficult. This was particularly problematic due to the visible Moiré patterns on the display, caused by optical interference between the light guide and the LCD’s grid-like circuit pattern.

Despite numerous attempts to find a solution, the team struggled to overcome this issue. Given the opportunity to tackle this challenge, it was a formidable task for me, especially as a new hire with a background in theoretical physics. Nonetheless, I developed a novel method to effectively eliminate the Moiré patterns, utilizing a specific irregular dot pattern for the light scatterers. The accompanying figure illustrates the stark contrast in uniformity between the conventional method and our innovative approach. An intriguing discovery was that random numbers defined mathematically don’t always appear random to the human eye. My solution leveraged mathematical theory to adjust the irregularity level and molecular dynamics simulation for precise control. For more details, please refer to the presentation slides.

This breakthrough was subsequently integrated into ThinkPad A30p, which was the world’s first laptop PC equipped with a UXGA IPS display.