Program

To accelerate element-wise calculation in arrays, we applied SIMD to parallelize array manipulation while developing numeric analysis program. However, SIMD intrinsic library contains lots of similar functions for different data types. We applied meta-programming to reduce duplicated codes cause by those similar functions in order to keep the maintainability. In this talk we will share our journey on meta-programming for SIMD from function-like C-style macro to C++ template programming and some template tips such as type matching, SFINAE.

BEVFormer is a transformer-based architecture for 3D object detection and multi-view perception. It projects multi-camera images into a bird’s-eye view (BEV) space and applies spatial and temporal attention to model the scene across time and viewpoints. This enables accurate and efficient object detection without relying on LiDAR data. In this talk, we will explore the core ideas behind BEVFormer, including its BEV representation, attention mechanisms, and how it integrates multi-camera inputs over time.

儘管大型語言模型(LLM)的持續進步，數學能力逐漸加強，在碰到新穎且需要邏輯推演的問題，LLM 仍表現的不可靠。在本計畫中，我們嘗試一個新的途徑，讓 LLM 以「操作外部工具」的方式來使用代數運算系統(computer algebra system, CAS)。