Research Topics
 |
Learning with Imperfect DataOur core research focuses on understanding and classifying complex, evolving systems, ranging from human behavior to industrial processes. Since real-world data is rarely perfect and can be unlabeled, weakly or noisy labeled, few, class imbalanced, or biased , our goal is to learn under these imperfect, multimodal scenarios. To tackle this challenge, we investigate flexible models that can perform under unseen conditions through domain adaptation and generalization , explore few-shot and zero-shot learning to learn from small amounts of data , and develop continual learning approaches to handle new incoming data in real-time. |
|
Learning with Imperfect DataOur core research focuses on understanding and classifying complex, evolving systems, ranging from human behavior to industrial processes. Since real-world data is rarely perfect and can be unlabeled, weakly or noisy labeled, few, class imbalanced, or biased , our goal is to learn under these imperfect, multimodal scenarios. To tackle this challenge, we investigate flexible models that can perform under unseen conditions through domain adaptation and generalization , explore few-shot and zero-shot learning to learn from small amounts of data , and develop continual learning approaches to handle new incoming data in real-time. |
Lightweight Adaptation of Foundation ModelsWhile Large Multimodal Models (LMMs) and Foundation Models (FMs) are currently state-of-the-art, they are generalist by nature and not tailored for specific tasks. In AIGO, we focus on the lightweight adaptation of these models by reusing the best "engine" and improving it with task-specific components. We utilize "Adapters"—small add-on modules that specialize FMs for new tasks and domains. This approach is highly parameter-efficient, requiring only a few extra parameters to be retrained , and composable, allowing multiple adapters to be stacked or swapped based on the application. Furthermore, this efficiency supports Edge AI, enabling the design of low-power AI modules deployable on embedded or portable devices. |
 |
|
Lightweight Adaptation of Foundation ModelsWhile Large Multimodal Models (LMMs) and Foundation Models (FMs) are currently state-of-the-art, they are generalist by nature and not tailored for specific tasks. In AIGO, we focus on the lightweight adaptation of these models by reusing the best "engine" and improving it with task-specific components. We utilize "Adapters"—small add-on modules that specialize FMs for new tasks and domains. This approach is highly parameter-efficient, requiring only a few extra parameters to be retrained , and composable, allowing multiple adapters to be stacked or swapped based on the application. Furthermore, this efficiency supports Edge AI, enabling the design of low-power AI modules deployable on embedded or portable devices. |
 |
Multimodal Data & Generative AITo comprehensively analyze complex environments, our group manages and integrates different data modalities, including visual (image, video, 3D, pose), audio, and language data. We place a strong emphasis on multimodal and embodied data representation learning. Beyond processing existing data, we employ generative models to synthesize new data for augmentation. Generating new data by conditioning multimodal Foundation Models helps us avoid or limit costly real-world data collection and annotation while allowing for the control and validation of the data generation process. |
|
Multimodal Data & Generative AITo comprehensively analyze complex environments, our group manages and integrates different data modalities, including visual (image, video, 3D, pose), audio, and language data. We place a strong emphasis on multimodal and embodied data representation learning. Beyond processing existing data, we employ generative models to synthesize new data for augmentation. Generating new data by conditioning multimodal Foundation Models helps us avoid or limit costly real-world data collection and annotation while allowing for the control and validation of the data generation process. |
Applications
Healthcare & Social AI
Aligning with the "AI for Good" vision, our algorithmic research is heavily applied to human-centric domains, including healthcare, assisted living, human-computer interaction, and robotics and social agents. We develop models capable of action recognition, human pose and motion understanding, and behavior analysis. In healthcare, we utilize adapters fine-tuned on radiology data, reports, and scan interpretations. For human monitoring, our adapters are tuned to interpret multimodal signals—such as video, audio, and physiological data—for activity or fatigue recognition.
Industrial Inspection & Aerospace
AIGO's expertise extends to developing intelligent systems for the industrial and aerospace sectors. In manufacturing, we train specialized adapters to detect defects or anomalies in industrial components. Our research also drives advancements in predictive maintenance and industrial monitoring, employing domain-specific adapters for sensor-based state estimation. Moreover, we leverage continual learning methods to ensure reliable, continual fault detection in evolving industrial environments , and apply our robust AI methodologies to tackle the unique challenges of the aerospace domain.