Molecular representation learning is an essential field focusing on understanding and predicting molecular properties through advanced computational models. It plays a significant role in drug discovery and material science, providing insights by analyzing molecular structures. The fundamental challenge in molecular representation learning involves efficiently capturing the intricate 3D structures of molecules, which are crucial for… →

Language models, a subset of artificial intelligence, focus on interpreting and generating human-like text. These models are integral to various applications, ranging from automated chatbots to advanced predictive text and language translation services. The ongoing challenge in this field is enhancing these models’ efficiency and performance, which involves refining their ability to process & understand… →

Large Language Models (LLMs) like GPT-3 and ChatGPT exhibit exceptional capabilities in complex reasoning tasks such as mathematical problem-solving and code generation, far surpassing standard supervised machine learning techniques. The key to unlocking these advanced reasoning abilities lies in the chain of thought (CoT), which refers to the ability of the model to generate intermediate… →

Autoregressive language models (ALMs) have proven their capability in machine translation, text generation, etc. However, these models pose challenges, including computational complexity and GPU memory usage. Despite great success in various applications, there is an urgent need to find a cost-effective way to serve these models. Moreover, the generative inference of large language models (LLMs)… →

Quantization, a method integral to computational linguistics, is essential for managing the vast computational demands of deploying large language models (LLMs). It simplifies data, thereby facilitating quicker computations and more efficient model performance. However, deploying LLMs is inherently complex due to their colossal size and the computational intensity required. Effective deployment strategies must balance performance,… →

Mixture-of-experts (MoE) architectures use sparse activation to initial the scaling of model sizes while preserving high training and inference efficiency. However, training the router network creates the challenge of optimizing a non-differentiable, discrete objective despite the efficient scaling by MoE models. Recently, an MoE architecture called SMEAR was introduced, which is fully non-differentiable and merges… →

Understanding and mitigating hallucinations in vision-language models (VLVMs) is an emerging field of research that addresses the generation of coherent but factually incorrect responses by these advanced AI systems. As VLVMs increasingly integrate text and visual inputs to generate responses, the accuracy of these outputs becomes crucial, especially in settings where precision is paramount, such… →

Maritime transportation has always been pivotal for global trade and travel, but navigating the vast and often unpredictable waters presents significant challenges. The advent of autonomous ships promises to revolutionize this domain, leveraging advanced sensors and Artificial Intelligence (AI) to enhance situational awareness and ensure safe navigation. The comprehensive integration of various sensor technologies with… →

The power of LLMs to generate coherent and contextually appropriate text is impressive and valuable. However, these models sometimes produce content that appears accurate but is incorrect or irrelevant—a problem known as “hallucination.” This issue can be particularly problematic in fields requiring high factual accuracy, such as medical or financial applications. Therefore, there’s a pressing… →