There are two major challenges in visual representation learning: the computational inefficiency of Vision Transformers (ViTs) and the limited capacity of Convolutional Neural Networks (CNNs) to capture global contextual information. ViTs suffer from quadratic computational complexity while excelling in fitting capabilities and international receptive field. On the other hand, CNNs offer scalability and linear complexity…

Many people are now interested in the state space model (SSM) because of how recent research has advanced. Modern SSMs, which derive from the classic state space model, benefit from concurrent training and excel at capturing long-range dependencies. Process sequence data across many activities and modalities using SSM-based methods like linear state-space layers (LSSL), structured…

Task-agnostic model pre-training is now the norm in Natural Language Processing, driven by the recent revolution in large language models (LLMs) like ChatGPT. These models showcase proficiency in tackling intricate reasoning tasks, adhering to instructions, and serving as the backbone for widely used AI assistants. Their success is attributed to a consistent enhancement in performance…

In the constantly evolving field of machine learning, particularly in semantic segmentation, the accurate estimation and validation of uncertainty have become increasingly vital. Despite numerous studies claiming advances in uncertainty methods, there remains a disconnection between theoretical development and practical application. Fundamental questions linger, such as whether it is feasible to separate data-related (aleatoric) and…

The practical deployment of multi-billion parameter neural rankers in real-world systems poses a significant challenge in information retrieval (IR). These advanced neural rankers demonstrate high effectiveness but are hampered by their substantial computational requirements for inference, making them impractical for production use. This dilemma poses a critical problem in IR, as it is necessary to…

The struggle to balance training efficiency with performance has become increasingly pronounced within computer vision. Traditional training methodologies, often reliant on expansive datasets, substantially burden computational resources, creating a notable barrier for researchers with limited access to high-powered computing infrastructures. This issue is compounded by the fact that many existing solutions, while reducing the sample…

A crucial area of interest is generating images from text, particularly focusing on preserving human identity accurately. This task demands high detail and fidelity, especially when dealing with human faces involving complex and nuanced semantics. While existing models adeptly handle general styles and objects, they often need to improve when producing images that maintain the…

Object detection plays a vital role in multi-modal understanding systems, where images are input into models to generate proposals aligned with text. This process is crucial for state-of-the-art models handling Open-Vocabulary Detection (OVD), Phrase Grounding (PG), and Referring Expression Comprehension (REC). OVD models are trained on base categories in zero-shot scenarios but must predict both…

MLLMs, or multimodal large language models, have been advancing lately. By incorporating images into large language models (LLMs) and harnessing the capabilities of LLMs, MLLMs demonstrate exceptional skill in tasks including visual question answering, instruction following, and image understanding. Studies have seen a significant flaw in these models despite their improvements; they still have some…

3D-aware Generative Adversarial Networks (GANs) have made remarkable advancements in generating multi-view-consistent images and 3D geometries from collections of 2D images through neural volume rendering. However, despite these advancements, a significant challenge has emerged due to the substantial memory and computational costs associated with dense sampling in volume rendering. This limitation has compelled 3D GANs…