This article reviews notable AI research papers published in Week 35 of 2024 (24W35), covering autonomous driving, efficient training, time series learning, 3D scene synthesis, motion generation, and AI-database integration.

Efficient Models: SAD (Spiking Autonomous Driving) proposes an energy-efficient autonomous driving system using Spiking Neural Networks (SNNs) — event-based computation enabling lower energy consumption while maintaining performance across perception, prediction, and planning modules on the nuScenes dataset. Inheritune enables efficient small LM pretraining by inheriting layers from larger LMs as initialization, achieving comparable performance with significantly less data and training time using GPT2 architectures on OpenWebText. TSDE (Time Series Diffusion Embedding) proposes self-supervised learning for time series via diffusion processes with IIF (Imputation-Interpolation-Forecasting) masks, achieving superior performance across diverse time series tasks.

3D/Motion Generation: Neural Assets introduces 3D-aware multi-object scene synthesis by extracting object visual representations combined with 3D pose information, enabling independent object control in generated scenes. T3M improves speech and text-driven human motion generation through improved alignment between multimodal inputs and body movements. MTMamba++ advances multi-task dense scene understanding through improved Mamba-based architecture handling multiple visual tasks simultaneously. TAG demonstrates Text-to-SQL enhanced with graph-based schema understanding, showing the potential for LLM-database integration in enterprise analytics applications.