|A few guidelines can keep your AI implementation from getting derailed.|
(Image source: Niek Verlaan from Pixabay)
In the blink of an eye, AI has gone from novelty to urgency. Tech leaders are telling companies they need to adopt AI now or be left behind. And a recent Gartner survey shows just that: AI adoption has skyrocketed over the last four years, with a 270 percent increase in the percentage of enterprises implementing AI during that period.
However, the same survey shows that 63 percent of organizations still haven’t implemented AI or machine learning (ML) in some form.
Why are there so many organizations falling behind the curve?
We meet with companies every week that are in some stage of their first ML project. And sadly, most of the conversations go more or less the same way. The project is strategic and highly visible within the organization. The internal proof of concept went off without a hitch. Now, the team is focused on getting the model’s level of confidence to a point where it can be put into production.
It’s at this point – the transition from proof of concept to production software development – that the project typically runs into big trouble. When we first meet with data science teams, their budget is often dwindling, their delivery deadline is imminent, and their model is still underperforming.
Sound familiar? The guidelines below might help your organization get its AI model to production on time without blowing your budget.
1.) Don’t Task Your A Team with Rookie Work
Newly-hired enterprise data scientists are always assured that the organization has all the source data they’ll ever need to build and test an AI system. What the scientists are not told is that the data is unusable. Rather than bringing AI to bear on a pressing public or corporate problem, these starry-eyed -- and very expensive -- data scientists instead find themselves cleaning up, organizing, and normalizing mountains of raw data.
These kinds of data preparation tasks are high on everyone’s list of recommended AI activities to outsource. Let your data team focus on disruptive and innovative work.
2.) Get with the Times; Get Agile
Enterprise software development teams learned a long time ago that agile development methodologies produce better software and do so faster than the traditional waterfall approach.
The waterfall method treats a big complex software system like a monolith that has to pass muster at each of a number of development stages before it can move to the next. The modern agile approach breaks a complex system into smaller discrete parts, each of which can independently navigate the stages of planning, coding and testing. The agile approach finds issues within the project earlier, which saves time and money.
3.) Don't Underestimate the Challenge of Training Data
With the proof of concept behind them, data science teams begin the real work of building out and training their algorithm. Few of them are prepared for the scale of the training data task. It’s not unusual for them to confront a 100x or even 1000x jump in training data required over the proof of concept phase of the project.
Data scientists understand very well how much data they require. This is part of their training. What they don’t grasp until they have an AI project under their belt is the scope of effort required to prepare that much training data.
The result is data scientists who are stuck in a seemingly endless slog of labeling and annotating data day and night, while making little measurable progress against the end goal.
Training data preparation is another activity that is frequently outsourced because it requires technologies, workforces, and project skills that most data science teams lack. But if you plan to keep this task in-house, ensure you bring in all the resources required to train your data. If you are not sure what volume of data you need, outsource this area of the project to defend your budget and team.
Finally, model training never stops, even after deployment. Therefore, your training apparatus – whether in-house or outsourced - must stay in place. Know that once your model has reached an appropriate confidence level, your work is still not complete.
Companies can get their AI and machine learning models to production faster and on-budget by setting clear expectations and responsibilities for its data science team, using the agile approach, and preparing for the challenge of training datasets. If the rate of companies implementing AI continues to skyrocket, a realistic view of what it takes to get a model to an appropriate confidence level will be vital. Don’t fall victim to wasted time and a blown budget.
Don Roedner is head of marketing for Alegion, a training-data platform for artificial intelligence (AI) and machine-learning initiatives. Don has over 25 years of experience working with B2B software companies as a marketer and, previously, in a variety of technical roles.
The nation's largest embedded systems conference is back with a new education program tailored to the needs of today's embedded systems professionals, connecting you to hundreds of software developers, hardware engineers, start-up visionaries, and industry pros across the space. Be inspired through hands-on training and education across five conference tracks. Plus, take part in technical tutorials delivered by top embedded systems professionals. Click here to register today!