Credit: Google News
Rip-and-replace your expensive hardware? Trust a startup? Or DIY? Here is real-world advice for easing into artificial intelligence and machine learning applications.
For all the hype about artificial intelligence (AI) and machine learning (ML), many IT managers are left scratching their heads about how to get started with these functions in their computer systems.
There’s often no good choice: Replace your expensive existing hardware with new systems from the likes of Dell and IBM, or put your faith in unproven small specialists who may not make clear what their software actually does.
SEE: Managing AI and ML in the enterprise (ZDNet special report) | Download the report as a PDF (TechRepublic)
Instead, independent experts say, it’s best to be slow and specific about embarking on AI or machine learning projects. Understand what the technology can really do, understand what problems you have to solve, and understand what it takes to make it happen. The first two parts are covered here, but what about the third aspect, which is implementation?
“The first thing you have to do is figure out what AI can actually do for you as a company,” said AI/ML consultant Adam Geitgey, who helps companies develop software and blogs extensively about it.
Presently, Geitgey said, where AI/ML software works best is in automating repetitive human tasks that require a small amount of judgment. “What you want to look for are places where you have a lot of people making decisions over and over… find something that is labor intensive that you do a lot of,” he explained.
Some examples include reviewing civil discovery documents in lawsuits, image classification, and transcribing audio. For internal IT functions, examples include tuning/optimizing your data center operations, configuration management, and systems patching/updating, analyst Henry Baltazar of 451 Research added.
SEE: Artificial intelligence: Trends, obstacles, and potential wins (Tech Pro Research)
Second, for all of these examples, “You need a lot of data to train AI to do that… if you don’t have that data you’re not going to be able to build an AI system,” he noted. You can buy off-the-shelf applications from Amazon, Google, and IBM, but if you need something custom you will have to assemble a team to build it.
“A lot of people hire specialists right now, but enough mid-level software developers are getting interested,” Geitgey observed. “It’s pretty immature first-generation stuff. You can imagine a couple of years out, these kind of tools will be much more available and standardized, and you probably won’t get them from your hardware vendor.” For now, “If it’s the first thing your company has ever done, it might be helpful to have guidance.”
“The third step then is actually creating the solution and testing the effectiveness,” Geitgey added. It’s common to read about hyperscale companies using AI for their internal computer maintenance operations, but that’s probably not efficient for normal-sized companies, he said.
Common mistakes include wanting to use AI/ML simply because it’s popular, and jumping into software development before understanding the problem you need to solve, Geitgey said.
SEE: Sensor’d enterprise: IoT, ML, and big data (ZDNet special report) | Download the report as a PDF (TechRepublic)
In order to start collecting enough data to understand the problems and to develop smart enough software, “What I always advise is to tell CEOs and decision makers that the data itself is an asset to your company… especially if it’s something no one else has,” he said. “You want something on the order of 10,000 data points to do something useful.” And make sure the data is relevant, Geitgey noted—online sales figures from summer won’t help your software predict the necessary compute cycles for Cyber Monday.
After building or buying AI/ML software, you also have to understand how to measure whether it delivers on the promises, 451’s Baltazar said. Currently less than half of developers understand how to do this, he said. Indicators to watch for include improved efficiency (such as fewer employees needed to perform the work), fewer IT trouble tickets, and faster remediation, he said.
Credit: Google News