Sara Hooker - The Hardware Lottery, Sparsity and Fairness
Machine Learning Street Talk (MLST)
00:00
Navigating the Costs and Complexities of Machine Learning
This chapter explores the rising costs associated with specialized hardware for machine learning, emphasizing the challenges of training advanced models like capsule networks and the limitations of current research methodologies. The discussion highlights the potential of sparse neural networks and model compression techniques as solutions to reduce costs while maintaining performance, advocating for a shift in focus towards optimizing model parameters. Additionally, it examines differing perspectives on model design versus computational power, underscoring the need for innovative architectural approaches to improve efficiency in deep learning.
Transcript
Play full episode
