Every IT problem has a learning curve and a tipping point, an aha! moment when the solution reveals itself. Working on AI projects, I experienced the same; the moment came when I realized how different the logic is in AI from traditional software programming. I had acquired the AI mindset that allowed me to really get started with AI.
AI is sometimes referred to as "software 2.0" because of the fundamentally different way it is programmed. Instead of telling the computer what to do, we must let the computer learn by itself. To make it "smarter," we must feed it with more and better training data. The AI then "programs itself," by detecting patterns in that data set.
Because of this, it is sometimes hard for us to know what the AI really learned. There have been cases where an AI that was trained to detect sheep actually learned to detect grass, or an AI trained to detect criminals by judging their faces was actually functioning as a smile detector. These are the extreme examples, but the less error-prone are actually worse as they carry our racial or gender bias into the AI's decision algorithm, where it is even harder to detect.
So, getting AI right is important. Here are three pieces of advice for how to get started with AI.
1. Set realistic expectations of what AI can and can't do for your business
AI can be good enough, even if it provides a 2% boost or is only 30% accurate. The question is not whether it's good enough; it is how to utilize an AI technology ... to give some kind of incremental boost.
Xun WangCTO, Bloomreach
To get started with AI, you first need to have reasonable expectations of what AI can do for your business.
Unfortunately, the media and vendor hype surrounding AI makes it hard to form reasonable expectations. Warnings about an AI apocalypse, for example, make the machines seem far more intelligent than they are, leading to misguided expectations, which, when not realized, cause the entire project to fail.
Most projects measure AI on its accuracy compared to how humans perform in the same situation. A neural network, for example, detected cancer with 94% accuracy, beating human radiologists. While the healthcare industry is particularly sensitive and needs super-accurate AI, the case does not hold for many other industries. A video recommendation engine, for instance, would do just fine with an 80% accuracy rate. Other projects could be fine with a 30% to 40% accuracy.
Xun Wang
"AI can be good enough, even if it provides a 2% boost or is only 30% accurate," claimed Xun Wang, CTO at Bloomreach, an AI-driven digital experience platform. "The question is not whether it's good enough; it is how to utilize an AI technology or component within a larger system to give some kind of incremental boost."
So, the measure is not AI's accuracy percentage, per se, but how much AI is helping your business. Instead of expecting the AI to completely replace your staff, the real measure should be how much it augments them, makes the job easier for them, increases their accuracy and frees up their time.
2. Find the value in company data
To get started with AI, you need data -- quality data. The need for good data -- and lots of it -- in AI may seem obvious, since this is a technology that requires to data to work. But it's surprising to me how easily the data collection step is missed.
I believe this is because the IT profession has not yet adopted an AI mindset. To cultivate an AI mindset requires CIOs to study the processes of their business, to discover the data they collect (or need to collect) and find ways to optimize or improve the business via AI accordingly.
This data can be sitting in your CRM, in your website analytics logs or support tickets. With the AI mindset, you'll start to think of this data differently. It won't be the contents of that data that matter, but the patterns and features that you can extract from it.
By "data" we don't mean only demographics or surfing behavior. AI is incredibly dumb in that it has no clue what it does, yet it's also extremely smart in detecting patterns we could not have thought of. What AI lacks in human understanding, it makes up for in speed of calculation.
For example, in the field of natural language understanding, researchers have fed AI with gigabytes of text. Their aim was to have the AI find similarities and relationships between words -- for example, the relationship between the words "woman" and "man" has a similar structure as that between a "queen" and a "king." In early July, scientists used this mechanism to study tons of scientific abstracts for discovering previously overlooked chemical compounds.
Once you get started with AI, you'll see every "data lake" as an AI optimization opportunity: You have the building blocks data; adding AI enables you to explore the data for untapped potential. For instance, one could develop a support system that would recommend solutions to customer or user problems before they post anything to the operators based on the context of what they are writing and not just on trigger keywords. Or through sentiment analysis, detect the severity of the customer issues and assign your senior staff to the more urgent cases accordingly.
Of course, there will be cases where you find the need for AI, but you don't have the necessary data. In those cases, you can either start to collect data internally, or acquire or purchase the necessary data from third-party sources to get started faster.
3. Start with simple algorithms and keep on track
When you are getting started with AI, it is important to think small.
Ruslan Gavrilyuk
"Start from the processes that have the least barriers in terms of overcoming inertia and the ones that allow for the best results tracking," recommended Ruslan Gavrilyuk, co-founder and president of market intelligence provider TeqAtlas. "Start small, automate one process at a time and proceed only after visible results."
One reason to start small relates back to the need to train an AI. Nothing is more frustrating than training an AI for hours and learning in the end that the results are unreliable. For that reason, simpler algorithms, which are trained much faster, offer significant help in showing if you are on the right track.
"Bear in mind that it takes time to gain visible results, let alone ROI," continued Gavrilyuk. Don't try to pull off deep learning or convolutional networks on your first run.
Similarly, your AI team must prepare the data, which is easier said than done. To use an oft-quoted maxim in the BI world: "garbage in, garbage out." It's easy to get advertisements or notes from sidebars when scraping webpages, which are completely irrelevant to the contents of the main page. Likewise, the quality of the data matters. When Open AI trained its "monstrous" GPT-2 model, it only relied on articles with high upvotes on Reddit.
Even during the training, your AI developers should have a look at how the AI model is understanding the data. A spam filtering engine, for instance, can show what words it most likely counts as spam and what it does not -- a quick look under the hood would reveal if the AI is on the right track.
Only after the data is clean and the AI model and the algorithm is fairly well understood should the AI developers attempt to optimize it by moving to more advanced algorithms such as convolutional neural nets or recurrent neural networks. And at that time, Gavrilyuk recommended that you "rent out computational power and storage, rather than buy and maintain it."
The time to get started with AI is now
According to research by TeqAtlas, people lack understanding of AI applications. Cultural barriers prevent companies from adopting AI, as many business leaders either consider the technology as their job-eliminating rival or lack trust in terms of handing over controls to machines.
But AI is not only very exciting, it is also very necessary, particularly as the competition ramps up. If you know what the requirements are to get started with AI, what you should reasonably expect and how to get there, you will know how to harness the power of this new technology and the excitement will translate into business results.
Dig Deeper on IT applications, infrastructure and operations
Xun Wang
Ruslan Gavrilyuk
