Preliminary System for Performance Measurement of Deep Learning Clusters; The Attention Economy: How Tech Companies Respond to Backlash 360 views

How can the efficiency of existing systems be improved? Efficiency problems apply to both computational systems and the human users of digital systems. Improvements in either system will produce a given output with limited waste of finite resources. As datasets and number of parameters for machine learning models grow larger, it becomes harder to train models on single systems, which lack memory and power. Training must be apportioned among several systems and then aggregated, incurring communication overhead among the systems. Because features and settings for different machine learning jobs vary, resource demands and bottlenecks vary too. The researchteam measured job completion time and resource utilization under several configurations, and studied theoretical optimal configurations. This new work will test configurations in non-ideal situations, where worker nodes become stragglers, or resource capacities are exceeded on purpose. In future work, researchers will develop techniques to relieve the bottlenecks for the various models. New technologies have brought convenience, but have also brought new distractions, harmful to user productivity and wellbeing. Tech companies practice limited self-regulation and other methods that controversial industries have used to dampen criticism and avert public regulation. Limited or ineffective self-regulation protects profitable business models at a cost to public interests.

BS (Bachelor of Science)

attention economy; tech company regulation; natural monopoly; deep learning clusters; distributed training; social media

School of Engineering and Applied Science Bachelor of Science in Computer Science Technical Advisor: Haiying Shen STS Advisor: Peter Norton

English

All rights reserved (no additional license for public reuse)

2021-05-12