For most companies, it doesn’t make sense to embrace AI as part of their core business. They should procure the expertise as a service, writes Albert van Breemen.
Data-driven artificial-intelligence approaches such as machine learning and deep learning have some major advantages over traditional engineering techniques. In areas such as computer vision, speech recognition and optimal control, data-driven AI algorithms achieve significantly higher performance levels. An additional benefit is that these algorithms adapt to the latest data available and thus are optimized to actual real-world situations. Moreover, when modeling an engineering problem with traditional techniques proves too difficult, a data-driven AI approach might provide an alternative way to solve the problem.
Despite these many advantages, real-world engineering applications with artificial intelligence are still rare. The problem with AI isn’t the theory but rather putting it into practice.
One obstacle when applying artificial intelligence to engineering problems is the lack of AI knowledge within traditional engineering teams. Data-driven approaches are so new that a solid understanding of what it takes to adopt AI and integrate it into an engineering system is lacking. Next to an iterative engineering process of defining a problem and finding a solution, AI introduces a new iterative experimentation lifecycle that needs to be managed. This lifecycle consists of collecting data, annotating data, training a model and deploying a model. A new way of working with AI skills and tools is required to be successful.
Another issue with putting AI into practice is the over-expectation of the technology’s capabilities. In general, training a model with an accuracy of up to 80 percent is easy and can be done quickly. The real challenge is getting above 95 percent, or even to 100 percent accuracy. This requires scaling from a few hundred to tens of thousands of data points, increasing the quality levels of data annotations, tweaking model architectures and more. In some cases, such accuracy levels can only be achieved by considering the application in total and redesigning sensor systems to collect data differently. Here, a 20/80 rule would apply: achieving the last 20 percent accuracy requires 80 percent of the effort.
A further challenge is the rapid pace at which the AI technology stack is changing. From AI hardware to deep-learning frameworks and model architectures, every few months, new versions and developments are released. Creating a stable AI technology stack and managing all the dependencies between the various components – on top of previously mentioned obstacles – is a real challenge that requires a specialist.
For some companies, it may be cost effective to hire such a specialist; for others, particularly smaller and non-tech ones, that option doesn’t seem as attractive. That’s where AI as a service comes in: external teams supporting customers in integrating AI into their applications while managing the AI technology stack and the deep-learning lifecycles.
Although most companies want to follow the AI buzz and add it to the core of their business, that would be inadvisable. Artificial intelligence is a complex and specialized technology that requires a serious investment in setting up a dedicated AI team that develops the right tools and platforms. Such an investment often is justified only when you want to become an AI technology provider. Do not fall into the ‘not invented here’ trap: companies should focus on their core business, and invest and innovate at the application level.