Data Analytics Newsletter #12
November 2019

,

The Data Analytics Practice Committee (DAPC), the Young Data Analytics Working Group (YDAWG) and the Institute are pleased to bring the latest in the world of Data Analytics to your inbox, and to share some of our recent work with you.

In this edition, we look at the GPT-2: 1.5B release and the use of AI in business, including ethics and potential disastrous consequences.

The UK Actuarial Post has an article on the popular buzz-word of “AI” - Is AI overhyped or an essential business tool? It highlights the importance of investing in systems in streamlining data collection and delivering better customer outcomes as well as better data quality, and complementing AI solutions with professional advice from (human) brokers.

Ethics in AI

However, having the correct ethical framework is important too. In the article Be careful what you wish for, Colin Priest highlights four key principles for AI ethics, and two case studies in health where data science failures have led to models that produced biased outcomes.

When AI goes wrong

Poor quality models can have disastrous consequences. The investigation into the fatal accident involving a self-driving Uber vehicle failed to recognise a pedestrian until it was too late.

The "known unknowns"

In Normal Deviance – Seeing the forest for the analytics trees, Hugh Miller highlights how analytics tend to optimise the “known universe” – but can be poor predictors outside of that. Data scientists need to consider the “known unknowns” and consider how those risks can be mitigated.

The need for interpretability in machine learning

How can models be trained, or used to produce more robust outcomes? Interpretability in models can help highlight potential biases within the model fitting process. In this article by Josh Jaroudy, he discusses the importance of finding causal relationships, the value in simplicity, and controlled experimentation.

GPT-2: 1.5B release

This month, OpenAI has finally released the full 1.5B parameter GPT-2 natural language model. The release was staged over concerns about its potential misuse in generating misleading news, social media posts, or other fraudulent activity.

Writing this Newsletter takes time each month, so would these developments allow us to finally automate it? Let us see with a writing prompt and three different runs of the model:

Prompt: "So, there are several reasons why we must conclude that actuaries make the best data scientists. Firstly,"

Here is the response: “They have a better understanding of how the world actually works than most people. Secondly, their training prepares them well for the task of data science. They are familiar with the statistical techniques and concepts that are important in the development of data science, and they can develop a sound understanding of the statistical issues that are the key to the success of data science in practice.
Finally, the fact that actuaries work with data often leads to a more holistic approach that leads to a more accurate understanding of the impact of their data, as well as the way to improve the quality of the data.

That was...surprisingly coherent. The prompt had to be deliberately worded to guide the response in a particular direction, some data quality artefacts gets replicated in the output, and the responses do veer off-topic eventually, but it made some fairly reasonable arguments of the quality you may see in a blog or newsletter.

You can test out the different writing prompts here.

Editor’s note

Calling all Melbourne members: Does data analytics excite you? Would you like to meet like-minded people, and make some events happen? I'd like to hear from you! Please get in touch!

Past editions of this Newsletter are now available here.

Disclaimer

The Institute wishes it to be understood that any opinions put forward in this publication are not necessarily those of the Institute.

Contents

When AI goes wrong

The "known unknowns"

GPT-2: 1.5B release