Tag: 2024
All the talks with the tag "2024".
An Introduction to Neural Differential Equations
Jyotirmaya ShivottamPublished: at 10:00 AMThis talk outlines the basics of Neural Differential Equations (NDEs), starting with the foundational paper by Chen et al. on Neural Ordinary Differential Equations (NODEs) and explores connections with dynamical modeling. It presents some examples of neural architectures in the NDE framework and discusses the universal approximation and expressivity properties of certain NDEs, in the context of the manifold hypothesis. The talk concludes with practical considerations for implementing NDE networks.
Communication Denied, Bee-Inspired Swarm Robotics for Search and Rescue
Jyothish Kumar JPublished: at 03:45 PMLi et. al.'s paper titled Swarm Robotics Search-and-Rescue (SAR) (2023) was discussed. The contributions of the paper, viz. grouping of target sites using a combination of mean shift and genetic algorithms, and swarm behaviour in an environment, where 2-way communication is denied, were understood.
Understanding the Intuition and Math Behind Stable Diffusion
Annada Prasad BeheraPublished: at 01:30 PMThis talk provides insights into Stable Diffusion, a state-of-the-art technique for generative models in image and video synthesis. It explains the intuitive and mathematical foundations of stable diffusion, offering a deeper understanding of how it works.
The Era of 1-bit LLMs - A Brief Overview
Adhilsha AnsadPublished: at 02:00 PMThis overview discusses the concept of 1-bit Large Language Models (LLMs) based on the paper "The Era of 1-bit LLMs - All Large Language Models are in 1.58 Bits". It presents BitNet b1.58, a 1-bit LLM variant that achieves competitive performance with full-precision Transformer LLMs while being more cost-effective in terms of latency, memory, throughput, and energy consumption. The overview highlights the potential of 1-bit LLMs in defining new scaling laws, training models, and designing hardware optimized for 1-bit LLMs.
Scalable MatMul-free Language Modeling
Sagar Prakash BaradPublished: at 02:00 PMThis talk presents a paper that proposes a scalable MatMul-free language model, challenging the assumption that matrix multiplications are essential for high-performing language models. The paper demonstrates that by using ternary weights and element-wise Hadamard products, MatMul operations can be completely removed from large language models while maintaining strong performance. The paper provides an optimized implementation of the MatMul-free language model, achieving significant reductions in memory usage and latency compared to conventional models.
SplaTAM - Splat, Track & Map 3D Gaussians for Dense RGB-D SLAM
Anubhav VishwakarmaPublished: at 02:00 PMThis paper introduces SplaTAM, an approach that leverages explicit volumetric representations, specifically 3D Gaussians, for dense simultaneous localization and mapping (SLAM) from a single RGB-D camera. SplaTAM achieves high-fidelity reconstruction and superior performance in camera pose estimation, map construction, and novel-view synthesis compared to existing methods. The implementation code for SplaTAM is available on GitHub.