Research Spotlights from Singapore

Sponsored by the Association for the Advancement of Artificial Intelligence
January 22 – January 25, 2025 | Singapore EXPO, Singapore

---

Data-Centric AI for Large Language Foundation Models

Bryan Low

In this talk, I will discuss the motivation and practical implications of two recent research efforts of my GLOW.AI research group, namely, (1) resource-efficient AI (specifically, how to learn with less data, and (2) AI to check on AI (specifically, data attribution and provenance, and evaluation methods for data removal and machine unlearning).

Tackling Hateful Video Content with Multimodal AI

Roy Lee

As online video platforms become central to digital communication, moderating hateful video content has emerged as a critical and complex challenge. Unlike text or image-based content, hateful videos often operate across multiple modalities, visuals, audio, and textual overlays, and carry subtle or culturally specific signals that evade detection. In this talk, I will present two recent studies that aim to advance the frontier of hateful video moderation through multimodal and cross-modal learning. First, I introduce MultiHateClip, a multilingual benchmark dataset comprising 2,000 annotated short videos from YouTube and Bilibili. It provides nuanced labels, hateful, offensive, and normal, along with annotations on target groups and contributing modalities, revealing cultural distinctions in how hate is conveyed across English and Chinese content. Second, I discuss a novel cross-modal transfer learning approach that leverages more easily accessible hateful meme datasets to enhance video moderation systems. This framework includes a human-in-the-loop re-annotation process to align meme and video labels, and demonstrates how vision-language models fine-tuned on re-annotated memes can outperform video-only baselines. Together, these works highlight the promise of data-efficient and culturally aware multimodal AI models for online harm detection and open new pathways for scalable and globally relevant video moderation strategies.

Foundation Models’ Adaptation: A Case Study on Remote Sensing Imagery

Sun Qianru

Large visual-language models, such as CLIP (Contrastive Language-Image Pretraining) and SD (Stable Diffusion), demonstrate remarkable performance in general image recognition and generation tasks. Their continual learning involves two folds: enhancing their performance with more natural images as input, and adapting them to specialized image domains. Our research targets the latter, using remote sensing (RS) imagery as a use case. RS, which relies on specialized satellites, presents challenges in image annotation and suffers from data scarcity and class imbalance, especially in special spectrums. Adapting models in this domain often leads to strong biases, where features of major classes overshadow those of minor classes. To address this, we recently introduced several solutions such as debLoRA–a generic training approach compatible with various low-rank model adaptation methods (like LoRA) to produce debiased features. In this talk, we will delve into its details and its subsequent works.

Where We Hit the Limits: Blue-Sky Directions for Large Language and Multimodal Models

Soujanya Poria

Despite the revolutionary impact of the Transformer architecture and the “Attention Is All You Need” paradigm, the field is currently facing a performance plateau characterized by diminishing returns on scaling and the computational bottlenecks of quadratic attention complexity . This talk explores “blue-sky” directions to overcome these limitations through innovations in model architectures and techniques. First, we address the instability issues inherent in current linear attention mechanisms, such as DeltaNet, which suffer from error accumulation due to Euler discretization . We introduce Error-Free Linear Attention (EFLA), a novel framework that derives an exact solution to the underlying continuous-time ordinary differential equation (ODE) . By leveraging Rank-1 matrix dynamics, EFLA achieves a computationally tractable, linear-time update rule that eliminates discretization error, offering guaranteed stability and superior performance in language modeling tasks. Second, we examine the intersection of vision-language reasoning and robotic control (VLA models), identifying the limitations of Supervised Fine-Tuning (SFT) and the scalability challenges of reinforcement learning in robotics . We present NORA-1.5, a VLA architecture that couples an auto-regressive backbone with a Flow-Matching Action Expert . By utilizing World Model-guided rewards and Direct Preference Optimization (DPO), NORA-1.5 enables scalable post-training on synthetic data, achieving state-of-the-art results on benchmarks like SimplerEnv and LIBERO while demonstrating robust real-world deployment on the Galaxea A1 robot . These contributions collectively chart a path toward more efficient, stable, and capable AI systems.