AI Meets AI: The Key to Actually Implementing AI
AI Meets AI: The Key to Actually Implementing AI
What you're about to see is a fundamental shift in how machine learning capabilities are delivered — and we aren't just talking deployment in the cloud versus on-premise.
Join the DZone community and get the full member experience.Join For Free
Insight for I&O leaders on deploying AIOps platforms to enhance performance monitoring today. Read the Guide.
With AI use case scenarios becoming more complex, the actual implementation of AI has also grown more challenging. However, a new way of delivering AI may be on the horizon.
An unprecedented amount of progress was made with AI and machine learning in 2017, as numerous companies deployed these technologies in real-world applications. This trend is projected to hold true through the near future, with some analysts like Gartner predicting that AI technologies will be in every new software product by 2020. From healthcare diagnosis to predictive maintenance for machines to conversational chatbots, there is no question that AI is quickly becoming a fundamental requirement for modern businesses.
However, despite the market buzz, many companies are still stumped by the prospect of deriving actual business value from the use of AI. In fact, the introduction of AI to actual products and solutions remains one of the leading sticking points for businesses, with many left asking, "How do I actually implement an AI solution?"
The Growing Complexity of AI Applications
As AI goes from a "nice to have" to a "need to have," it's also evolving in terms of complexity. Companies need more than just simple, standardized AI services that do image or text recognition — they need complex predictive scenarios that are highly specific to their operations and customized for their business needs.
For example, take a scenario that uses time series data to generate business insights, such as predictive maintenance for the Industrial Internet of Things (IIoT) or customer churn analysis for a customer experience organization. These scenarios can't be supported by simply calling a generic service with a few specific parameters and getting a result. Getting accurate and actionable results in these predictive scenarios requires a lot of data science work, with data being used over time to iteratively train the models and improve the accuracy and quality of the output. Additionally, businesses are being challenged to engineer new features, run and test many different models, and determine the right mix of models to provide the most accurate result — and that's just to determine what needs to be implemented in a production environment.
Moreover, businesses need to realize that AI is no longer the exclusive domain of data scientists and the engineers that help prepare data. The situation is not unlike how digital transformation has branched out from being an IT-driven initiative to a company-wide effort. Organizations must move beyond a siloed AI approach that divides the analytics team and the app development team. App developers need to become more knowledgeable about the data science lifecycle and app designers need to think about how predictive insights can drive the application experience.
To be successful, organizations must identify an approach that enables them to easily put models into production in a language that is appropriate for runtime-without rewriting the analytical model. Organizations need to not only optimize their initial models but also feed data and events back to the production model so that it can be continuously improved upon.
This may seem like a big, complicated process, but it's key to the actual implementation of AI — the AI of AI, if you will. AI will become unreachable to your organization if you cannot do this.
The New World of AI
So how can organizations effectively implement AI in a way that enables them to address complex predictive scenarios with limited data science resources? And how do organizations achieve success without retraining their entire development team?
The truth of the matter is that it can't be done by simply creating a narrowly defined, one-size-fits-all approach that will get you results with only a few parameters. It requires a more complex implementation to be insightful, actionable, and valuable to the business.
Take, for example, an IIoT predictive maintenance application that analyzes three months of time series data from sensors on hundreds or thousands of machines and returns the results automatically. This isn't a simple predictive result set that is returned, but a complete set of detected anomalies that occurred over that time, with prioritized results to eliminate the alert storms that previously made it impossible to operationalize the results. These prioritized results are served up via a work order on a mobile app to the appropriate regional field service personnel, who are then able to perform the necessary maintenance to maximize machine performance. It's a complex process where the machine learning is automated and feature engineering is done in an unsupervised fashion. The provided results analyze individual sensor data, machine-level data, and machine population data and are packaged up in a format that enables the business to take action.
Welcome to the new world of AI implementation. While it's a very new concept, the best market definition of this process is currently "anomaly detection." But not all solutions take the same approach and not all solutions deliver predictions that lead to better business outcomes. What you are about to see is a fundamental shift in how machine learning capabilities are delivered — and we aren't just talking deployment in the cloud versus on-premise. We are talking about a shift from delivering data science tools that make the data scientists more effective to data science results that eliminate the need for the data scientist to have these tools in the first place. In this brave new world, data scientists would be able to spend their time analyzing and improving the results, instead of wasting their time on non-mission-critical tasks.
The only thing that is required is that the data is provided in a time series format. Otherwise, you simply upload the data to the cloud (but on-premise options will exist too) and the automated AI does the rest, with accurate results returned within days.
Soon you can move from dreams of AI to actual implementation!
Published at DZone with permission of Mark Troester , DZone MVB. See the original article here.
Opinions expressed by DZone contributors are their own.