Reward functions play a crucial role in reinforcement learning (RL) systems, but their design presents significant challenges in balancing task definition simplicity with optimization effectiveness. The conventional approach of using binary rewards offers a straightforward task definition but creates optimization difficulties due to sparse learning signals. While intrinsic rewards have emerged as a solution to… →

Training large-scale AI models such as transformers and language models have become an indispensable yet highly demanding process in AI. With billions of parameters, these models offer groundbreaking capabilities but come at a steep cost in terms of computational power, memory, and energy consumption. For example, OpenAI’s GPT-3 comprises 175 billion parameters and requires weeks… →

Breaking down videos into smaller, meaningful parts for vision models remains challenging, particularly for long videos. Vision models rely on these smaller parts, called tokens, to process and understand video data, but creating these tokens efficiently is difficult. While recent tools achieve better video compression than older methods, they struggle to handle large video datasets… →

Semantic segmentation of the glottal area from high-speed videoendoscopic (HSV) sequences presents a critical challenge in laryngeal imaging. The field faces a significant shortage of high-quality, annotated datasets for training robust segmentation models. Therefore, the development of automatic segmentation technologies is hindered by this limitation and the creation of diagnostic tools such as Facilitative Playbacks… →

Graph Neural Networks have emerged as a transformative force in many real-life applications, from corporate finance risk management to local traffic prediction. Thereby, there is no gainsaying that much research has been centered around GNNs for a long time. A significant limitation of the current study, however, is its data dependency—with a focus on supervised… →

Neural machine translation (NMT) is a sophisticated branch of natural language processing that automates text conversion between languages using machine learning models. Over the years, it has become an indispensable tool for global communication, with applications spanning diverse areas such as technical document translation and digital content localization. Despite its advancements in translating straightforward text,… →

Natural Language Generation (NLG) is a domain of artificial intelligence that seeks to enable machines to produce human-like text. By leveraging advancements in deep learning, researchers aim to develop systems capable of generating contextually relevant and coherent responses. Applications of this technology span diverse areas, including automated customer support, creative writing, and real-time language translation,… →

FineWeb2 significantly advances multilingual pretraining datasets, covering over 1000 languages with high-quality data. The dataset uses approximately 8 terabytes of compressed text data and contains nearly 3 trillion words, sourced from 96 CommonCrawl snapshots between 2013 and 2024. Processed using the datatrove library, FineWeb2 demonstrates superior performance compared to established datasets like CC-100, mC4, CulturaX,… →

Multimodal reasoning—the ability to process and integrate information from diverse data sources such as text, images, and video—remains a demanding area of research in artificial intelligence (AI). Despite advancements, many models still struggle with contextually accurate and efficient cross-modal understanding. These challenges often stem from limitations in scale, narrowly focused datasets, and restricted access to… →