Module 14 2022

01/04/2022

Labelling large datasets

The Organisation for Professionals in Regulatory Affairs

19

So what’s the problem?

• Very large accurately labelled datasets are required for some applications and may be difficult to obtain. • Training networks is very computationally expensive. • Wrong answers can be very hard to explain, and perfect accuracy is not possible. This presents challenges in risk management.

2008, $110 Million, ~2.3MW “Roadrunner” supercomputer

2020, £318,000, ~3kW Nvidia NGX-A100

The Organisation for Professionals in Regulatory Affairs

20

10