Demystifying AI

Demystifying AIAsk any IT executives about the hottest enterprise IT topic and the sexiest job in 2017, the answers are like to be artificial intelligence (AI) and data scientist.  But according to technology providers and academics, AI is nothing new and data scientists have become a less sexy job.

The sci-fi concept of machine mimicking human’s cognitive functions was an academic discipline since 1956. But in recent years it is starting to realize, to some degree, in the business environment. According to IDC, machine learning (ML) and AI will produce 30-40% cost savings and productivity improvements in operations management by 2020.

In Asia, more CIOs are also looking to adopt AI. Gartner’s CIO Agenda survey indicated 37% of Asia Pacific IT leaders noted they have deployed or are in short-term planning to deploy AI for its business environment.

Machines do not learn

Despite the enthusiasm and recognition in the market, executives from analytic software provider SAS noted that the popularity of AI simply reflects the advancement in data modeling and processor technologies.

“AI is based on sets of models and machines don’t really learn,” said Jim Goodnight, CEO and co-founder of SAS. “People that write about machine is learning are caught up with the [concept of] neural networks, which is basically applying models.”

At the company’s recent user conference SAS Analytics Experience 2017 in Amsterdam, Goodnight said analytics is about fitting data in models and AI is simply using new types of modeling to provide predictions.

“AI is the computerization of what human could do,” added SAS CTO Oliver Schabenberger. “Algorithm helps to do the job better when problem is too big for us. We consider AI is just an extension of what we’ve been doing in analytics for the past 41 years.”

The new types of modeling and extension of analytics in AI that Goodnight and Schabenberger referred to is the application of artificial neural networks (AAN), which are computing systems inspired by biological neural networks, to develop deep learning algorithms.

These algorithms enable systems to progressively improve performance (i.e. learning). With the use of ANN, the system is able to improve faster and better than using the traditional computing algorithm using rule-based programming.

Schabenberger added that deep learning allows SAS to reexamine its existing analytics offerings like sentiment analysis with “new lenses.”

“AI technology today based mostly on deep learning, so we have heavily invested in that area and we’ll release our deep learning solutions and tool kit at the end of this year with support of GPU,” he said.

Specialized to process image, graphic processing unit (GPU) has extended from processing images in online games to enabling deep learning neural networks. Multiple studies, including a research from Hong Kong Baptist University, have indicated GPUs are achieving significant speed over CPUs in processing deep learning neural networks. The development in computing power is also another major reason for the popularity of AI.

Analytics 4.0

In addition to the development in data modeling and processors technologies, the rise of AI is also a reflection of autonomous analytics in Analytics 4.0, according to academics, authors and consultant Tom Davenport.

The Harvard Business Review author previously categorized the evolution of analytics into three stages. Analytics 1.0 is the era of business intelligence, when data from different systems were recorded, aggregated and analyzed. It is followed by Analytics 2.0, the era of big data, when data from external sources like the internet and social media is contributing to the analysis. It is compared to small data that is purely generated from a firm’s internal transaction systems. Analytics 3.0 is when organizations are developing data-oriented offerings and services, creating a data economy.

Davenport noted that in the era of Analytics 4.0, these data-enriched offerings among individual firms are connected to become pervasive and automated networks.

“Goods and services, traffic, communications, energy, money—all of these will flow around the network in massive volumes and at unprecedented speed,” he said. “No humans need apply to run these networks, since they couldn’t keep up with the activity or make decisions rapidly enough to help.”

Autonomous analytics take over sexy job

To enable such automation, Davenport noted autonomous analytics play a key role.

He explained throughout history, analytics is about applying different algorithms to a set of data to derive results. This process of designing algorithm, choosing the right model and identifying the missing data has been the job of the data scientist, which Davenport ones named it the sexiest job of the 21st century.  

But in Analytics 4.0, this process will be replaced by machine. “Machine can help to identify the right algorithm,” he said. “The models are going to be created, in some degree of, automation and the consumptions of it, not by human, but by machines to make decisions and take actions.”

With autonomous analytics, the computing systems can choose the best models, identify models that are violating any assumptions, and take actions accordingly. “This is having an interesting effect on organizations and it is making the data scientist job a little less sexy,” he said.

Davenport added that we are still at the early stage of Analytics 4.0, as a lot of fundamental issues in Analytics 1.0 and Analytics 2.0 remained to be solved.

“We don’t really have good data standards across the domain for IoT,” he said.  “We can’t fit all data into one standard and thus not a whole lot of automation yet.”

DELTA to success

He noted that the success in AI, similar to any analytics initiatives, relies on data, enterprise, leadership, target and analyst (DELTA).

In his book Analytics at Work: Smarter Decisions, Better Results, Davenport stated that clean, common and integrated data is the prerequisite for anything analytical. The enterprise’s approach and leadership commitment to manage data to avoid fiefdoms of data are other key components in the formula. Having a specific purpose or strategic target to apply the analytical models and a team of analysts are the last two of the five fundamentals.

Davenport said that simply applying some of these simple fundamentals with automation can reveal significant tangible benefits for enterprises. He quoted GE’s data integration exercise, which took a few months by using machines to integrate data of its suppliers across the entire group. It allowed GE to benefit with US$80 million saving.

Economics of AI

From the economist’s perspective, “AI simply lower the cost for human to make predictions,” said Ajay Agrawal, AI economist and founder of the Creative Destruction Lab at University of Toronto.

Agrawal said AI technologies use analytical models and computing power to calculate different predictions faster and cheaper than human. “The machine is replacing human to make predictions. The value of human prediction falls,” he said.

Predictions and judgement are two complements to make decisions. By lowering the cost to make predictions, AI raises the demand for human judgment.  “Machines make predictions, they don’t have judgement, so human judgement gets more valuable,” he said.

Agrawal noted that although AI appears to be powerful to provide predictions, it is merely a tool to bring a substantial input for making decisions—prediction is not decision making. Putting this theory into business context, AI is simply a tool to raise productivity for executives to making decisions for the business. 

He advised business executives “to start a business strategy before building the AI tools,” as machines do not learn and they can never make decisions without human judgement.

 

Image from iStockphoto

Read more on