The growth of data in the digital age presents both opportunities and challenges. An immense volume of text, images, audio, and video is generated daily across platforms. Traditional machine learning models, while effective in many scenarios, often struggle to process high-dimensional and unstructured data without extensive preprocessing and feature engineering. This approach is not only…
Generative Large Multimodal Models (LMMs), such as LLaVA and Qwen-VL, excel in vision-language (VL) tasks like image captioning and visual question answering (VQA). However, these models face challenges when applied to foundational discriminative VL tasks, such as image classification or multiple-choice VQA, which require discrete label predictions. The primary obstacle is the difficulty in extracting…
Large Language Models (LLMs) and Vision-Language Models (VLMs) transform natural language understanding, multimodal integration, and complex reasoning tasks. Yet, one critical limitation remains: current models cannot efficiently handle extremely large contexts. This challenge has prompted researchers to explore new methods and architectures to improve these models’ scalability, efficiency, and performance. Existing models typically support token…
Advances in large language and multimodal speech-text models have laid a foundation for seamless, real-time, natural, and human-like voice interactions. Achieving this requires systems to process speech content, emotional tones, and audio cues while giving accurate and coherent responses. However, challenges remain in overcoming differences in speech and text sequences, limited pre-training for speech tasks…
Scientific metadata in research literature holds immense significance, as highlighted by flourishing research in scientometrics—a discipline dedicated to analyzing scholarly literature. Metadata improves the findability and accessibility of scientific documents by indexing and linking papers in a massive graph. Today, the research community has realized the importance of metadata. However, its awareness and consideration were…
Speech processing systems often struggle to deliver clear audio in noisy environments. This challenge impacts applications such as hearing aids, automatic speech recognition (ASR), and speaker verification. Conventional single-channel speech enhancement (SE) systems use neural network architectures like LSTMs, CNNs, and GANs, but they are not without limitations. For instance, attention-based models such as Conformers,…
Biometric authentication has emerged as a promising solution to enhance security by offering a more robust defense against cyber threats. However, hackers can increasingly develop sophisticated methods to bypass traditional security measures as technology advances. This includes forging common protections such as easily guessed PINs, passwords, or even misplacing physical keys, which were once considered…
Blockchain systems face significant challenges in efficiently managing and updating state storage due to high write amplification (WA) and extensive I/O operations. In traditional architecture, such as Merkle Patricia Tries (MPT), frequent and expensive disk interactions incur inefficiencies that restrict throughput and scalability. Such problems are one of the biggest bottlenecks for decentralized applications requiring…
Mathematical reasoning has long been a significant challenge for Large Language Models (LLMs). Errors in intermediate reasoning steps can undermine both the accuracy and reliability of final outputs, which is particularly problematic for applications requiring precision, such as education and scientific computation. Traditional evaluation methods, like the Best-of-N (BoN) strategy, often fail to capture the…