SAS chief points the way to successful AI implementation

Artificial intelligence (AI), machine learning and digital transformation are some of the most talked about topics in industry, government and society today. And while a lot of talk surrounding them is hype, SAS Chief Operating Officer and Chief Technology Officer Dr Oliver Schabenberger asserts that these technologies are very real and powerful. Here are some of the key takeaways from his keynote at SAS’s 2019 Road2AI event.

“AI implementation requires a formal approach”:

According to Schabenberger, by way of Gartner, three out of four senior managers consider data science to deliver significant value and be essential to their organisation’s success. But in reality, most organisations are in the very early stages of their machine learning and AI journey.

In order to boost AI use, organisations need to take a formal approach. Baidu Vice President and Chief Scientist Andrew Ng lays out one such approach in a Harvard Business Review article.

In it, he suggests starting with one or two small, short-term, industry-specific projects which are most likely to create value for the company. This will likely give you a quick win and help drive AI adoption.

Automation does not equal autonomy:

Today, AI and machine learning users can choose their level of automation — from tuning a model to building the entire thing. But that doesn’t mean users can just ‘set and forget’.

In order to be successful, users need to feed algorithms with the right type, quantity and quality of data. This is especially important as businesses start to rely more heavily on unstable behavioural data, rather than demographic data.

Additionally, users need to monitor algorithms’ outputs and provide feedback in order to improve algorithms’ accuracy and efficiency.

“AI is not fundamentally biased”:

 AI, machine learning and analytics are not inherently biased, but the data people use and the models they build can be.

Importantly, Schabenberger explained that eliminating bias isn’t as simple as removing race, gender, sexual orientation and other attributes from the data because these variables are correlated with others.

“A model that contains a proxy for gender still contains gender information and the decisioning system can see these protected attributes and the fingerprints live on,” he said.

To combat this, users need to understand their data, how bias and discrimination are entering a decisioning system and what they expect a ‘successful’ model to look like.

Additionally, users should work towards developing ‘minimum virtuous products’ — offerings that test for the effects on stakeholders and build in guards against potential harms — rather than minimum viable products, according to Schabenberger, quoting General Catalyst Managing Director Hemant Taneja.

Here, SAS can provide tools for model interpretability and natural language generation so that ethicists and decision-makers can understand what the results actually mean.

Used correctly, AI and machine learning can provide valuable insights, help decision-making processes and make jobs easier and more productive. With a formal AI implementation strategy and a strong understanding of data and analytics, these technologies can move from being ‘hype’ to being real and powerful.