Tag: 2024
All the talks with the tag "2024".
Vision Mamba - A Brief Overview
Adhilsha AnsadPublished: at 02:00 PMThis talk presents a look into a variant of the Mamba architecture in vision tasks, a linear-time sequence modeling approach that uses selective state spaces to achieve results comparable to or even better than state-of-the-art vision models.
A Brief Overview of Recent Advances in Sequence Modeling Through Structured State Space Models
Jyotirmaya ShivottamPublished: at 10:00 AMThis talk outlines recent advances in sequence modeling using structured state space models, starting with the basic formulation that initiated this line of research and moving on to more recent works (Mamba) that present complete architectures operating under certain assumptions. Core techniques that make these models work efficiently on long sequences, such as discretization and an associative scan operation to parallelize the computation, will also be discussed.
Quaternion Graph Neural Networks
Rucha Bhalchandra JoshiPublished: at 02:00 PMRecently, graph neural networks (GNNs) have become an important and active research direction in deep learning. This talk proposes Quaternion Graph Neural Networks (QGNN) to learn graph representations within the Quaternion space, a hyper-complex vector space. The talk covers state-of-the-art results on benchmark datasets for graph classification and node classification, as well as knowledge graph completion.
Mamba - Linear-Time Sequence Modeling with Selective State Spaces
Sagar Prakash BaradPublished: at 02:00 PMThis talk presents a look into the Mamba architecture, a linear-time sequence modeling approach that uses selective state spaces to achieve state-of-the-art performance on a range of sequence modeling tasks.
Nitty Gritty Details of the CUDA Rasterizer Used in Gaussian Splatting
Annada Prasad BeheraPublished: at 02:00 PMThis talk delves into the details of the CUDA rasterizer used in Gaussian splatting. It covers the theory behind Gaussian splatting, the implementation of the `cuda_rasterizer.cu`, and the backward pass that enables optimizing algorithms such as Adam or SGD to optimize the Gaussians with respect to a given input image and produce novel views of the scene.