In 2021, you’d be hard pressed to find a company that isn’t using AI. AI has become more accessible and ubiquitous in recent years, and the term has cemented itself as a core component of the technological lexicon. What’s not part of the vocabulary — yet is increasingly critical to successful AI deployments — is the concept of composite AI.
Last summer, Gartner included composite AI as one of its five new innovation profiles in the annual Hype Cycle for Artificial Intelligence, defining the term as “the combination of different AI techniques to achieve the best results.” It does this by synergizing a mix of different types of AI — machine learning, traditional rules-based systems, optimization techniques, natural language processing and graph techniques — in order to improve AI systems’ learning efficiency, level of “common sense” and ability to solve a wide range of business problems.
While it may not sound especially earth-shattering because the concepts behind composite AI are ultimately not new, techniques from different AI domains offer unique approaches to address different aspects of an overall business problem. Using them in concert and not relying on a single one can have a powerful multiplier effect. With composite AI, organizations can create solutions that explore and exploit all aspects of the knowledge embedded in the data. That’s a truth we’ve seen come to bear for retailers, doctors and bankers — innovators across sectors that are taking advantage of composite AI today.
Composite AI in Action
We often hear businesses asking, “Which algorithm or AI technique should I use?” As AI tackles ever more complex problems, however, the best answer is often a combination of multiple techniques and technologies.
Consider a retail organization, for instance. Using composite AI, it could optimize its pricing and promotional efforts by gathering and cumulatively analyzing data from each store’s transactions, existing pricing and promotions, inventory levels, customer attributes and competitor pricing. Retailers could better understand pricing elasticity (i.e. the impact of any proposed pricing changes) for a subset of stores and/or items by concurrently leveraging machine learning techniques, and repeatedly testing and iterating based on customers’ reactions to any pricing modifications.
In medicine, Amsterdam University Medical Center has demonstrated the power of composite AI by combining technologies like computer vision, data visualization and machine learning to evaluate the efficacy of treatments for cancer patients. This mix enables doctors to visualize chemotherapy response based on the change in total tumor volume — an analysis that’s difficult to see with a naked eye, but can significantly improve overall survival rates.
And in the financial sector, the application of composite AI techniques is a game changer to the bottom line. By transitioning manual compliance reviews to an AI-powered, automated risk process, organizations are using natural language processing to extract crucial data from trade documents, conducting hundreds of compliance checks in real-time and using a gradient boosting model to automate processes and reduce false positives. Automation has enabled leading financial institutions to reduce compliance checks from 40 minutes to one minute — translating to millions of dollars in savings.
How to Get Started
The most important step in successfully leveraging composite AI is to first clearly define the business problem that requires solving. Once a company’s data science and AI practitioners have a deep understanding of the business problem at hand, they should be able to determine what available data sets will best inform the answer to the business problem. Selecting, integrating, and implementing the most productive combination of AI techniques typically comes back to the amount and type of data available.
For example, if a company’s most pressing business problem relates to customer experience and support, and textual or voice data is involved, two AI techniques — computer vision and natural language processing — should be in the mix of ingredients in the AI strategy. If a company is struggling to solve a financial problem within a specific department that involves primarily structured data, the composite AI recipe might call for a combination of statistics, machine learning and forecasting techniques.
After selecting the most productive combination of AI techniques, companies’ data science and AI practitioners can use machine learning pipelines to create models. To operationalize the resulting models and capture value as soon as possible, models or even decision flows can be embedded into relevant processes, devices, sensors or databases. Ensuring seamless and secure integration across every AI technique within a composite AI strategy is particularly critical, especially if open source libraries are being used and/or the company in question is within a highly regulated industry.
A Gradual, Integrated Approach is Key
Ultimately, composite AI is a team sport because it requires multidisciplinary proficiency and participation from experts across the entire spectrum of AI techniques. Rather than succumbing to overwhelm or focusing too much on one particular AI technique, companies should gradually build their composite AI strategy based on the business problem(s) they most want to solve.
Realistically, the mix of AI will also evolve over time as the business needs change and grow. It’s important to ensure the process is iterative and that a wide range of AI techniques are readily available. By integrating the most relevant combination of AI techniques and corresponding experts to solve a specific problem, and by continuously deploying models and decision flows to generate value, companies can reap the benefits of composite AI and realistically solve their most complex business challenges.
Brett Wujek, PhD, is principal product manager for analytics at SAS, the global leader in analytics. With more than two decades of experience in engineering, software development and data science, Brett helps direct advanced analytics development at SAS, particularly in the areas of machine learning and data mining. His formal background is in design optimization methodologies, receiving his PhD from the University of Notre Dame for his work developing efficient algorithms for multidisciplinary design optimization.
Source link
