4 Ways To Ensure AI Credibility And Adoption
I recently caught up with Jack Berkowitz, vice president of products and data science for Oracle Adaptive Intelligence, whose job it is to make sure the technology doesn’t end up as shelfware, but is put to good use for the business. He says he sees AI and machine elarning are now being embedded into a range of applications and functions, from supply chain and manufacturing applications to ERP, finance, procurement, human capital management, and customer experience applications for sales, service, marketing and commerce. The impact is nothing less than transformative, he adds, as some AI or machine learning-driven applications now in use today “are helping to unearth rich business insight and create greater efficiencies across the entire organization.”
Driven by these hopes of transformation to the positive side, the pace of artificial intelligence and machine learning adoption is accelerating. “AI is possible now because of the plethora of data, sophisticated algorithms, and lightning-fast computing power,” Berkowitz says. At the same time, he cautions that just because the technology is there, people will come to it. He provides some advice to drive acceptance of AI and machine learning:
Gradually introduce features — even small ones — helpful to end-users. Berkowitz says his own products at Oracle, for example, are designed “to send users subtle cues — slide-outs and alerts — to let them know of new AI features within the familiar user interface of the application.” He predicts other vendors will follow suit.
Ensure the right data is being applied to the right business problem. Along with user comfort with AI, there’s another concern that needs to be addressed: having good, quality data from the right sources. “The data that is needed for AI and machine learning is determined by the problem being solved,” Berkowitz explains. “For example, if you’re trying to identify customers to target a digital ad, then you can break this problem into different areas. First, how do I find out who are my most trusted customers, can I generate a customer importance score using their past purchases, lifetime value, or their activities on social networks? The next area is what data do I have about the campaign, such as metadata and target demographics. Once you have the data, then you can use techniques such as machine learning, deep learning and learning to rank to find and rank the customers most suitable for this ad campaign.”
Make sure data is of the best quality, and is double-checked. “The challenge starts with bringing in the fixed and transient data from various disparate sources into a common platform,” says Berkowitz. “Not only do you need to bring in the data, but you need to make sure the data is clean, changes are synchronized, and records denormalized. Maintaining the data quality and having checks at the data ingestion stage is critical.” He illustrates where things could go awry without checks on what data is being fed into the system: “Let’s say you’re bringing in ages into the training dataset. You normally have ranges between 23 to 68, but after the pipeline was pushed to production there was a bug in the data collection and now you have values at 220. Having checks in place would prevent the models from tuning its weights to these extreme cases.”
Test diligently. “Exploratory data analysis and feature engineering” is an important part of AI and machine learning success, Berkowitz says. “These approaches play a major role in driving important key features and insight that are not available from the data in its raw form. Various steps of feature scaling and feature transformations need to be done before the data can be passed into the algorithms. One of the key things the model does is generalize the patterns from past data in order to predict patterns in future data points.
