It’s no surprise that the market for machine learning (ML) as-a-service is set to explode over the next few years. Amazon, Google, Microsoft and IBM are all building out their cloud ML services, exposing them as APIs to make it easy for developers and data scientists to build, train and deploy new models.
The ability to harness algorithms via an API will be a massive leap forward in the adoption of intelligent applications for both enterprises and service providers. And that does not count what we have come to expect over the past few years: all this amazing capability is available as-a-service. How quickly we forget what a sea change it is to pay only for what you use. In this case, it means companies can pay only for the number of images they want classified, or the number of words they want spoken to a customer, for example.
But there is an interesting twist to this. Over the past few months, several service providers have told me they are finding more success rolling out their own models. And this is based on some really smart folks doing some serious benchmarks against the big cloud providers.
This shift appears to be happening mostly in the computer vision arena. For example, insurance companies are looking at ways to use machine learning to classify photos of accidents. Healthcare providers want to use ML to identify subtle changes in imaging scans. And energy companies want to use it to analyze video taken from drones to identify potential weaknesses in their pipelines. All these are very specific use cases that require accuracy and precision. So some providers have determined that in order to meet these requirements, they need to build the capability themselves. In the case of image recognition, which often means building a neural network, some really smart people will be doing some really complicated math.
To be clear, I have not heard of providers abandoning the massive-scale cloud vendors’ offerings in total – only for specific use cases that require a high degree of accuracy and precision. This makes sense: the models from the big guys must appeal to a very broad set of users across a lot of different use cases. But service providers are different. They are in the business of building bespoke solutions for specific needs.
But providers have another rationale as well. Amazon Rekognition, Google Cloud Vision, Microsoft Computer Vision and Watson Visual Recognition are all new offerings, so the support these vendors are providing for their APIs is “embryonic” at best. This will improve over time, but in the meantime providers are unwilling to bet their clients’ results on an API that may be finicky and not well supported.
Provider-built models could become sourcing differentiators as enterprises evaluate solutions. For example, if a provider can quantifiably benchmark that their claim resolution solution can classify and tag accident images better than AWS Rekognition, that will likely be a recipe for differentiation and growth.
It’s important to keep in mind that the key to all these models is the quality of the training data. The cloud providers have bigger scale than service providers (and probably better engineering talent), but in domain-specific problems like these, they don’t have better training data. That means the big cloud providers’ models can’t catch up to the proprietary ones developed by providers. It’s the enterprises that hold the treasure trove of transactions, images and voice conversations that can be used to train and optimize these models. So, whoever is closest to the enterprise – and therefore whoever is closest to the data – will be in the best position to build models specially tuned to these domain-specific challenges.
Measurement, Machine Learning, and Determinism
About the author
Stanton Jones helps clients maximize the value of their emerging technology investments. His current research focuses on the application of Intelligent Automation to enterprise operations, helping clients navigate the fast-moving ecosystem of technology vendors and service providers. Stanton is a recognized expert, and has been quoted in CIO, Forbes and The Times of London.