DNN Framework
Deep neural network (DNN) frameworks are the foundation of modern artificial intelligence, aiming to improve model accuracy, efficiency, and robustness. Current research focuses on optimizing DNN training and inference, including techniques like efficient parallelization strategies, mixed-precision training, and adaptive model selection based on resource constraints and carbon footprint. These advancements are crucial for deploying DNNs on resource-limited devices (e.g., edge computing) and mitigating challenges like adversarial attacks, noise, and data scarcity, ultimately impacting various fields from computer vision to natural language processing.
Papers
Harmony: Overcoming the Hurdles of GPU Memory Capacity to Train Massive DNN Models on Commodity Servers
Youjie Li, Amar Phanishayee, Derek Murray, Jakub Tarnawski, Nam Sung Kim
Auto-Transfer: Learning to Route Transferrable Representations
Keerthiram Murugesan, Vijay Sadashivaiah, Ronny Luss, Karthikeyan Shanmugam, Pin-Yu Chen, Amit Dhurandhar