AI as a Platform-as-a-Service
As AI matures, be on the lookout for AI service platforms, including some that will be able to develop a strategy on the fly for specific problems.
- By Brian J. Dooley
- June 29, 2017
The future of artificial intelligence (AI) is unfolding quickly and it doesn't always fit our current models. AI-as-a-service has been promoted by some as a category for chatbots and digital assistants, and by others as a platform characteristic. The reality is that it is likely to be both but with an emphasis on the latter. It's complicated.
AI Is a Component and a Method
On the surface, AI-as-a-service is a predictable development, but AI is not as coherent as other services. It tends to be a component, hidden in the background to provide extended capabilities, optimization, and user interface improvements. The major platform providers are simply integrating AI capabilities into their other platform offerings.
Platform providers with products in this area include IBM with Watson and its Bluemix platform; Amazon with Amazon Machine Learning; Google's network of AI offerings, including TensorFlow and its Cloud Prediction API; Microsoft with its Azure Machine Learning Studio; and others focusing on niches and vertical industries such as Salesforce's Einstein.
Platforms such as these provide a mix of development-oriented services that can be customized, integrated, and combined to create special offerings. The reason for this patchwork is that AI is more a set of algorithms and methods than a monolithic service. AI solutions must be strongly focused on specific applications. Complex solutions demand combinations of discrete AI services that may include basic "horizontal" skills (such as vision) and specific "vertical" skills (such as price comparison).
Horizontal skills include general applications such as natural language processing and image recognition, but these are not necessarily usable on their own. Vertical AI skills must be carefully crafted to meet the needs of their specific use cases and have a much narrower scope of application. A complete solution is likely to demand both types of skills, plus access to specific data, training, tweaking, and integration with complementary services.
The Move to Microservices
Meanwhile, software development itself is moving toward microservices and APIs to improve efficiency and mobility. For AI, this is critical because intelligence is a composite process. For example, autonomous vehicles incorporate horizontal AI services for vision and sensing, vertical AI processes for situational awareness, and other AI services for operation and navigation. These services are linked to additional analytics and data services to create a complex AI system.
A composite AI built of microservices makes sense, but this has a number of market implications. AI microservices need to be assembled, tested, and orchestrated for specific cases, but this goes against the user-friendly requirements of bringing AI "to the masses." The result is a set of offerings designed to manage AI APIs and microservices.
These services are available from the major platforms as well as from start-ups such as MLJAR.com, DataRobot (datarobot.com), PurePredictive (purepredictive.com), and Yhat (yhat.com). Unsurprisingly, some of these are themselves based on machine learning and AI, making it possible for the AI system to select algorithms and models and fit them to a particular use case.
Integrated Systems Use AI to Select Solutions
AI-as-a-service was first proposed as a category for cloud-based chatbots and digital assistants. Although these products are narrower in scope than what we think of as a complex, full AI offering, they can provide the entry point for an integrated cognitive response. This is already happening with platform offerings such as IBM Watson and Amazon Alexa.
Voice recognition and natural language processing are critical to the user interface, but the more interesting services being offered are microservice AI integration tools and services that assemble complete AI solutions to solve specific problems. For example, Salesforce.com's Einstein attempts to understand requirements from spoken requests and produces a result from a CRM cloud solution. IBM's Watson attempts to understand what numbers and analysis a user might be requesting and assembles a result from its analytics solutions.
With the proliferation of more advanced integrated solutions in platforms that are linked to other powerful capabilities, we can expect something more closely approaching human thought. A part of what we really define as intelligence is the ability to determine how to make a decision. We decide whether the answer lies in art or in science, if it is a calculation or a list, and so forth. The same applies to AI: Will this algorithm aid the current inquiry? Is the data available? Are other solutions available?
As we move forward into composite AI, we will move from general platforms-as-a-service and from concepts of AI-as-a-service into a middle ground of AI service platforms that will be able to develop a strategy on the fly for specific problems. We have not yet reached this point, but it appears inevitable.