Generalization & Adaptation under Distribution Shift

Mon, 01 Jan 0001 00:00:00 +0000

Developing machine-learning methods that generalize beyond their training distributions and adapt efficiently to new domains, tasks, and environments with limited supervision.

Research spans (ICML 2012); , enabling zero-shot transfer and compositional generalization to unseen domains (ICLR 2022); and for generalization to unseen target classes (ICLR 2022).

Recent work extends these ideas to real-world physical sensing systems, including and validating the resulting methods through field experiments under different sensing conditions (JLT 2024), as well as to downstream tasks under severe domain shifts with CLAP-S (ICASSP 2025).

, ICML 2012
, ICLR 2022
, ICLR 2022
, JLT 2024
, ICASSP 2025

Efficient Foundation-Model Adaptation & Post-Training

Mon, 01 Jan 0001 00:00:00 +0000

Modern post-training increasingly treats pretrained and specialized models themselves as reusable sources of knowledge, alongside data. My research explores efficient and reliable ways to adapt, compress, and integrate this knowledge for domain specialization and capability consolidation.

Research includes (NeurIPS 2024) and (NeurIPS 2025 Spotlight), which explore reparameterization and parameter sharing for extreme parameter efficiency in fine-tuning and their effects on optimization; (CVPR 2026), an uncertainty-aware knowledge distillation approach for reliably transferring knowledge from imperfect supervision into compact models; and combining capabilities from multiple specialized models into a unified multitask model.

, NeurIPS 2024
, NeurIPS 2025 Spotlight
, CVPR 2026

Small Multimodal Models: From Recognition to Reasoning

Mon, 01 Jan 0001 00:00:00 +0000

Many real-world deployments require small models, while scaling trends favor larger ones. How far can we push the capability frontier at small-model scale?

Developing small multimodal models for real-world settings where compute, memory, latency, privacy, or large-scale parallel processing make model capacity a first-class constraint. The broader goal is to advance small multimodal models from open-vocabulary recognition to multimodal understanding and reasoning.

Research progresses from memory-augmented few-shot adaptation that combines explicit knowledge from support-set exemplars with implicit knowledge encoded in model and adapter parameters ( , ICASSP 2025), to noise-robust multimodal learning by distilling a pretrained teacher into students trained under clean and noisy input conditions and adaptively fusing their representations through minimum-entropy selection, allowing both interpolation and extrapolation beyond the training conditions ( , ICASSP 2026), and toward small multimodal foundation models for domain-grounded question answering and reasoning.

, ICASSP 2025
, ICASSP 2026

Sensing AI for Real-World Physical Systems

Mon, 01 Jan 0001 00:00:00 +0000

Machine Learning for Optical Fiber Sensing over Deployed Telecommunication Infrastructure

Overview

Telecommunication fiber networks provide a large-scale sensing substrate that extends beneath cities and along long-haul routes. Distributed fiber sensing turns existing optical cables into sensors for the physical world, while modern machine learning makes it possible to extract useful information from the resulting large-scale spatiotemporal data.

Telecom
Networks

→

Distributed Fiber
Sensing

→

Physical Signals
at Scale

→

Machine
Learning

→

Physical-World
Intelligence

My research develops machine learning methods for distributed acoustic and temperature sensing, with an emphasis on turning raw sensing signals into reliable physical-world intelligence. Applications include infrastructure monitoring, cable protection, traffic and vehicle sensing, manhole localization and condition diagnostics, and other forms of situational awareness over deployed telecommunication networks.

The work spans learning from limited supervision, generalization and adaptation across sensing environments, probabilistic and generative modeling, foundation-model adaptation, and structured prediction. A central goal is to connect advances in machine learning with sensing systems that operate beyond controlled laboratory settings, using deployed infrastructure as both a source of real-world data and a testbed for generalization, adaptation, and validation.

Research Themes

Learning from Limited Supervision

Deployed fiber-optic sensing networks generate massive streams of spatiotemporal data, but obtaining precise task-specific annotations is often expensive and operationally difficult. My research explores how to turn this abundance of unlabeled and weakly labeled field data into an advantage, using weakly supervised learning, self-supervised learning, semi-supervised learning, and generative semi-supervised learning to learn useful sensing representations and build reliable models with limited manual labeling.

Generalization Across Sensing Environments

Models trained in one sensing environment or deployment configuration may not transfer reliably to another. My research studies inductive biases, representation learning, and adaptation mechanisms that improve generalization across sensing environments, routes, and configurations, with field deployments providing realistic tests of robustness.

Adaptation and Multimodal Foundation Models

Pretrained audio and multimodal models provide transferable knowledge for downstream sensing domains, but substantial domain shifts can arise when adapting them to fiber-optic sensing data. My recent work explores support-set adaptation, knowledge distillation, and noise-aware representation learning to efficiently transfer this knowledge to domain-specific sensing tasks.

Structured Prediction for Linear Infrastructure

Many critical infrastructure systems—including roads, highways, railways, pipelines, power lines, and telecommunication networks—extend over long linear corridors, making distributed fiber sensing a natural platform for large-scale monitoring. My research explores structured prediction along these corridors, using dense spatial information from distributed sensing together with geospatial context to continuously localize events and monitor infrastructure conditions in the physical world.

Research Progression

Representative work progresses from (Optics Express 2023), to that leverages unlabeled DAS data through semi-supervised modeling while supporting uncertainty calibration (ECOC 2025 Top-Scored Paper), to that uses sensing location as a pseudo-label to learn transferable representations across deployed fiber routes (JOCN 2026).

The research also progresses from (JLT 2024) to (JLT 2026; OFC 2025 Top-Scored Paper), leveraging spatial and temporal context for fine-grained localization and infrastructure condition monitoring. Across these efforts, methods are developed and validated through field experiments on existing telecommunication fiber networks.

Field Deployment & Real-World Validation

Field deployment has been an integral part of my sensing research, using existing telecommunication networks to develop and validate sensing-AI methods under real-world operational conditions.

Cable Protection → Infrastructure Diagnostics → City-Scale Monitoring

Cable Protection & Event Monitoring

Long-term abnormal-activity monitoring and threat assessment over deployed telecommunication fiber.

Representative field events include a fallen utility pole affecting aerial cable and construction activity near buried cable.

Infrastructure Localization & Diagnostics

Distributed acoustic and temperature sensing for infrastructure localization and condition assessment across deployed fiber routes.

Field trials in Richardson, TX, and Long Beach Island, NJ, demonstrate complementary DAS and DTS information for manhole localization and condition diagnostics.

City-Scale Multimodal Sensing

Long-term fiber sensing integrated with satellite imagery and AI for live telecommunication infrastructure management.

A 21-month field deployment over approximately 40 km of fiber, with construction activity cross-referenced using satellite imagery.

Datasets & Code

Public resources associated with this research include:

CLAP-S — code for support-set adaptation of pretrained audio-language models.
Gunshot–Firework Dataset — fiber-optic acoustic sensing dataset for downstream recognition.

Patents & Technology Transfer

Sensing-AI research has contributed to patented technologies spanning the full pipeline from machine-learning methodology to physical-world applications and operational systems:

ML Methodology for Fiber Sensing — methods spanning context encoders, self-supervised, contrastive, and weakly supervised learning; zero-shot domain adaptation and support-set adaptation of foundation models; generative and synthetic-data modeling; and memory-augmented learning for privacy-preserving system design.
Physical Infrastructure Sensing & Understanding — methods for localization, identification, and condition monitoring of buried and aerial cables, manholes and handholes, utility poles and fiber coils, roads and highways, vehicles and construction equipment, and other physical assets and landmarks along deployed fiber networks.
Real-World Sensing Applications — applications spanning cable protection and intrusion detection, traffic and vehicle sensing, infrastructure health monitoring, earthquake and acoustic-source localization, fiber-route mapping and geolocation, and public-safety use cases such as gunshot detection.
Systems, Prototypes & Technology Transfer — research prototypes integrate fiber sensing, AI, and real-time processing into end-to-end systems for field validation and operational use. Related NEC Labs America research has also progressed into commercial systems: the provides a modular AI-analytics platform over deployed fiber for applications including Fiber Cable Monitoring, Fiber Cable Identification, Fiber Cable Position Locator, and Pigtail Finder.

Tutorials & Talks

— Invited Tutorial, OFS30 2026, Raleigh, NC, USA
— Invited Workshop Talk & Panelist, OECC/PSC 2025, Sapporo, Japan

Research Highlights & Media

— NEC Labs research story on turning deployed telecom fiber into large-scale sensing networks through advanced fiber sensing and self-supervised AI.
— NEC Labs research story on compact, noise-aware audio-language models for real-world sensing and edge deployment.
AI/fiber-optic combo poised to improve telecommunications — Laser Focus World coverage of our fiber-sensing AI research: ·

Fiber Sensing Publications

2026

Wataru Kohno, Shaobo Han, Noriyuki Tonami, Tingfeng Li, Jingchen Sun, and Ting Wang. “ .” IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2026), Barcelona, Spain, 2026.
Shaobo Han, Ming-Fang Huang, Yue-Kai Huang, and Philip N. Ji. “Rapid State-of-Polarization Change Point Detection via Minimum Lossy Coding Length.” OECC 2026. Oral.
Scott Kotrla, Ming-Fang Huang, Jian Fang, Shaobo Han, Jamie Lynn, Ezra Ip, Jeffrey A. Mundt, and Ting Wang. “ .” Journal of Optical Communications and Networking, Vol. 18, No. 4, pp. B72–B84, 2026.
Shaobo Han, Ming-Fang Huang, Yaowen Li, Glenn A. Wellbrock, Tiejun J. Xia, Scott Kotrla, Jeffrey A. Mundt, James M. Moore, Philip Ji, Tingfeng Li, Yuheng Chen, Ting Wang, and Yoshiaki Aono. “ .” IEEE Journal of Lightwave Technology, Vol. 44, No. 3, pp. 1086–1093, 2026.

2025

Jingchen Sun, Shaobo Han, Wataru Kohno, and Changyou Chen. “ .” IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2025), Hyderabad, India, 2025. Lecture.
Wataru Kohno, Noriyuki Tonami, Jian Fang, Shaobo Han, Jingchen Sun, and Ting Wang. “ .” IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2025), Hyderabad, India, 2025.
Shaobo Han, Ming-Fang Huang, Tingfeng Li, Scott Kotrla, Jeffrey A. Mundt, Ting Wang, and Yoshiaki Aono. “ .” ECOC 2025. Top-Scored Paper.
Ming-Fang Huang, Tingfeng Li, Shaobo Han, Biplob Debnath, Eric Cosatto, Tina Zheng, Scott Kotrla, Glenn A. Wellbrock, Tiejun J. Xia, Jeffrey A. Mundt, Ting Wang, Yoshiaki Aono, and Koji Asahi. “ .” OECC 2025, PDP-A-4. Postdeadline Paper.
Fatih Yaman, Andrea D’Amico, Shaobo Han, Shinsuke Fujisawa, Eduardo Mateo, and Takanori Inoue. “ .” OECC 2025, WC3-3. Oral.
Shaobo Han, Philip N. Ji, and Ting Wang. “ .” OFC 2025, W1D. Oral.
Ming-Fang Huang, Shaobo Han, Yaowen Li, Glenn A. Wellbrock, Tiejun J. Xia, Scott Kotrla, James M. Moore, Philip Ji, Tingfeng Li, Yuheng Chen, Ting Wang, and Yoshiaki Aono. “ .” OFC 2025, M1C. Top-Scored Paper.
Tingfeng Li, Ming-Fang Huang, Shaobo Han, Yaowen Li, Glenn A. Wellbrock, Tiejun J. Xia, Scott Kotrla, James M. Moore, and Ting Wang. “ .” OFC 2025, M1C. Oral.

2024

Shaobo Han, Ming-Fang Huang, Tingfeng Li, Jian Fang, Zhuocheng Jiang, and Ting Wang. “ .” IEEE Journal of Lightwave Technology, Vol. 42, No. 12, 2024.
Zhuocheng Jiang, Yangmin Ding, Junhui Zhao, Yue Tian, Shaobo Han, Sarper Ozharar, Ting Wang, and James M. Moore. “ .” 20th IEEE Workshop on Perception Beyond the Visible Spectrum (PBVS), 2024.

2023

Glenn A. Wellbrock, Tiejun J. Xia, Ming-Fang Huang, Shaobo Han, Yuheng Chen, Ting Wang, and Yoshiaki Aono. “ .” IEEE Journal of Lightwave Technology, Vol. 41, No. 12, 2023.
Alexander Bukharin, Shaobo Han, Yuheng Chen, Ming-Fang Huang, Yue-Kai Huang, Yao Xie, and Ting Wang. “ .” Optics Express, Vol. 31, No. 6, pp. 9591–9607, 2023.
Fatih Yaman, Yang Li, Shaobo Han, Takanori Inoue, Eduardo Mateo, and Yoshihisa Inada. “ .” OFC 2023, W1J.7. Oral.

2022

Ezra Ip, Fabien Ravet, Hugo Martins, Ming-Fang Huang, Tatsuya Okamoto, Shaobo Han, Chaitanya Narisetty, Jian Fang, Yue-Kai Huang, Milad Salemi, Etienne Rochat, Fabien Briffod, Alexandre Goy, Maria del Rosario Fernández-Ruiz, and Miguel González Herráez. “ .” Proceedings of the IEEE, 2022.
Ming-Fang Huang, Jian Fang, Shaobo Han, Zhuocheng Jiang, Sarper Ozharar, Yuheng Chen, Tomoyuki Hino, and Ting Wang. “ .” OECC 2022, PDP-A-1. Postdeadline Paper.
Ting Wang, Ming-Fang Huang, Shaobo Han, and Chaitanya Narisetty. “ .” OFC 2022, Th3G.1.

2021

Ming-Fang Huang, Shaobo Han, Glenn A. Wellbrock, Tiejun J. Xia, Chaitanya Narisetty, Milad Salemi, Yuheng Chen, James M. Moore, Philip N. Ji, Giovanni Milione, Ting Wang, Yukihide Yoda, Yoshinori Kitahara, Morio Ito, Yoshiaki Aono, and Atsuo Itoh. “ .” OECC 2021, T5A.8. Postdeadline Paper.
Tingfeng Li, Yuheng Chen, Ming-Fang Huang, Shaobo Han, and Ting Wang. “ .” OFC 2021, Th1A.26.
You Lu, Yue Tian, Shaobo Han, Eric Cosatto, Sarper Ozharar, and Yangmin Ding. “ .” IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2021), Toronto, Ontario, Canada, 2021.
Tiejun J. Xia, Glenn A. Wellbrock, Ming-Fang Huang, Shaobo Han, Yuheng Chen, Milad Salemi, Philip N. Ji, Ting Wang, and Yoshiaki Aono. “ .” OFC 2021, Th4H.3. Oral. Tingye Li Innovation Prize Finalist.

Research Programs |

Generalization & Adaptation under Distribution Shift

Related Publications

Efficient Foundation-Model Adaptation & Post-Training

Related Publications

Small Multimodal Models: From Recognition to Reasoning

Related Publications

Sensing AI for Real-World Physical Systems

Overview

Research Themes

Research Progression

Field Deployment & Real-World Validation

Cable Protection & Event Monitoring

Infrastructure Localization & Diagnostics

City-Scale Multimodal Sensing

Datasets & Code

Patents & Technology Transfer

Tutorials & Talks

Research Highlights & Media

Fiber Sensing Publications

2026

2025

2024

2023

2022

2021