AI’s Rise Raises Control, Bias Issues: SAS Institute


The deluge of data is leading to the rapid adoption of artificial intelligence, but developments in AI technology could outstrip humans' ability to understand and control the potential impacts of that technology.

That's the word from SAS Institute Chief Technology Officer Oliver Schabenberger, who told an audience of IT executives and users at this week's Nvidia GPU Technology Conference that AI is real but over-hyped.

Schabenberger's comments came just days after the Cary, N.C.-based analytics software developer announced it was investing $1 billion in AI with a focus on R&D innovation, education initiatives focused on addressing the importance of understanding and benefiting from AI, and expert services to help customers optimize returns on their AI projects.

Schabenberger defined artificial intelligence as a non-biological system that performs tasks or makes decisions that humans can do.

Sponsored post

[Related: 6 Boldest Statements From Jensen Huang And Eyal Waldman On Nvidia-Mellanox Deal]

Therefore, he said, the rise of AI is not surprising. "Technology wants to be organized," he said.

The growth of AI stems from two primary drivers, Schabenberger said.

The first is digital transformation, which is all about the diffusion and democratization of data, he said. The world is moving from hardware to software, with books, cars, and hospitals becoming bits and bytes. "And the result is an explosion of data," he said.

The second is changes in connectivity from physical roads and trails and money to digital interactions between people as well as IoT, which is connecting everything, Schabenberger said.

When the two are combined, the result is massive volumes of structured and unstructured data flowing through the Internet, Schabenberger said. "All of a sudden, data has latency," he said. "And if you are late with data, you miss opportunities."

Schabenberger said there are two types of AI.

The first is what he termed "narrow AI," which is what is available today and which is continually improving, often without human intervention. He cited GPT-2, a project of OpenAI, which was trained to predict the next word in 40 GBs of Internet text given all the previous word within the text.

However, he said, GPT-2 went beyond what it was taught to actually learn to count, make lists, and even do rudimentary text translation.

"Narrow AI keeps improving," he said.

The second type is what he termed "artificial general intelligence," or AGI, which are thinking machines. "Not purpose built, but can perform any task, like us. … [But we have no clue how to build AGI," he said.

The first wave of AI started decades ago as the rules-based automating of tasks, Schabenberger said. "We were codifying logic," he said. "Expert systems like that, we have been building a long time."

However, he said, that traditional AI technology cannot handle abstract data, and cannot perceive the world around it. Those are being addressed by the current wave of AI that uses large amounts of data and predictive learning. "When exposed to a new set of data, it can predict without having been told what to do," he said.

Other parts of AI, including deep learning and neural networks, are also not new, but are now becoming more powerful thanks to massive networking power combined with big data, as well as new algorithms, Schabenberger said.

However, he said, deep learning is not a solution. "It's a hammer, but the real world does not look like nails," he said.

He cited as an example tax return software, which today does not use deep learning but instead returns results based on the data inputted. Going forward, he said, deep learning will allow the software to actually understand the tax code and make decisions based on it.

Deep learning, however, is not the right way to do tax returns, given that it would be based in part on past tax returns, which would also include fraudulently-based returns, he said. "I don't want my tax returns based on other peoples' fraudulent taxes," he said.

Furthermore, he said, he might want to know how his taxes were estimated, and would want his taxes based on his own circumstances, and not that of others.

The tax return example shows the danger of relying on patterns to make decisions, Schabenberger said. Without patterns, AI may offer incorrect results. "And bias in the system might impact how a decision is made," he said.

The good news about deep learning is that it is only one subset of AI, Schabenberger said. "We have many tools," he said. "We should look at all the toolsets we have."

Building AI requires a lot of understanding about what is being built and about what data is applied, and bias can easily creep into the AI model, Schabenberger said. "If we understand how bias impacts the algorithms, we can reverse it," he said.

But that is not easy, Schabenberger said. Sex, race, gender are all biases, and it is important to understand when to use them and when not to. He cited the example of a medical AI system. "If you look at diseases, some diseases are more common in women," he said. "So that bias is still in there."

Schabenberger said a joke in the software industry states that all software works as coded. "But the important thing is, does software work as designed?" he said.

With AI, that changes as humans are no longer in charge of the coding, he said. "And we also don't work on how the data is used," he said.

Schabenberger said businesses are in a rush to adopt AI because of a fear they may be left behind. However, he said, they struggle with incoherent strategies, and one in three organizations are held back by a lack of skills. "These issues are also holding back companies from implementing data strategies," he said.

He suggested that companies start small and identify a handful of issues they can do, such as a speech-based response system. However, he said, businesses should only build their own if such systems are industry-specific, while taking advantage of other developers' non-industry-specific systems when possible. "Focus on value in your industry sector," he said.

Businesses should also accelerate their AI projects with third-party support, appoint a business leader to see the project through, and start with a pilot project and small team to build the skills needed.

Also, he said, it is important that any AI projects address at least one of three requirements. "It has to reduce costs, increase revenue, or solve a particular issue," he said.