Machine Learning Development Platform Race Heats up as AWS Launches SageMaker

AWS already hosts an array of third-party machine learning (ML) service platforms for developing AI algorithms, and even had its own Amazon Machine Learning service. However, the market perception of Amazon Machine Learning was that it barely scratched the surface of what was possible with such a platform, and that other offerings from Google, Microsoft, IBM, and new entrant Algorithmia were more capable. ML service platforms bridge the gap between AI frameworks and developers, allowing developers and data scientists to research and produce ML models in a user-friendly environment, a layer above framewoks like TensorFlow and Caffe2.

At AWS re:Invent, Amazon launched a new AI development platform, SageMaker, stepping up its ML service offering and targeting companies looking for an easy entry point into developing ML applications. Amazon claims that “SageMaker is a fully managed end-to-end machine learning service that enables data scientists, developers, and machine learning experts to quickly build, train, and host machine learning models at scale.” Amazon handles all the underlying infrastructure required to run an ML model, including any issues like node failure, auto scaling, or security patching. The focus of the SageMaker platform is accessibility, which should prove attractive to companies unwilling to undertake the steep and costly learning curve associated with developing ML applications.

The market for ML service platforms like SageMaker is a competitive space, with some of the largest tech players battling it out. Google already has a number of ML service frameworks, including TensorFX, Google Prediction API, and Google ML engine. Microsoft has Azure Machine Learning Studio and IBM, of course, has Watson. Amazon already had Amazon ML, although the platform was far less sophisticated and functional than SageMaker.

SageMaker is compatible with the major AI software frameworks and third-party libraries can be imported into the platform. Algorithms from MXNet, TensorFlow, Gluon, Caffe2, PyTorch, Keras, Microsoft Cognitive Toolkit, and Torch can all be ported onto SageMaker’s platform. For all this supposed flexibility, AWS is focused on pushing SageMaker’s 10 custom built-in modules. In contrast, other platforms like Azure Machine Learning Studio offers in excess of 100 built-in algorithms, while Algorithmia’s AI Layer service contains a library of more than 4,000 free and purchasable algorithms. For a developer unfamiliar with artificial intelligence (AI) software development, the sheer number of options available on other platforms is not necessarily helpful; in streamlining options for developers, the product could prove attractive to developers and enterprises looking to develop a model quickly. AWS has already released a list of customers for SageMaker, including Thomas Reuters, Hotels.com, bookkeeping software provider Intuit, CRM software vendor Zendesk, and satellite company DigitalGlobe.

According to AWS, SageMaker significantly reduces the education and work required to develop an AI application. For instance, SageMaker’s hyperparameter optimization feature eases the transition from “research-grade” ML models to “production-grade” models. This process ordinarily involves parameter tuning, with engineers assessing which parameter changes affect the outcomes of the model and then tuning the model accordingly. SageMaker’s hyperparameter optimization feature replaces that process, by automatically spinning out multiple copies of a research model and then using ML to asses which parameter changes affect the performance of the ML model itself. However, competitors have pointed out that SageMaker is a compute-intensive platform, which translates into higher costs for its users. Larger entities that already have developers working on ML are less likely to be attracted to the SageMaker platform, particularly if they have a preference over who is in control of their AI ecosystem at a higher level.

Undoubtedly, Amazon has identified a key market opportunity. There are many easy wins for companies looking to implement some of the basic data analytics made possible by ML models. A lot of greenfield market space exists, which could lower the technical bar for AI, enabling AWS to access the market with SageMaker. Many entities not versed in ML will now have an easy entry point for ML software development. ABI Research spoke with Algorithmia, which said it was particularly impressed by the package offered in SageMaker and it anticipated that the platform would become the market leader. If the market transpires as Algorithmia anticipates, this could be a significant embarrassment for IBM Watson and Microsoft, which made entries into this segment earlier than Amazon. AWS has already pulled in a lot of enterprise customers for its cloud services and these should also act as a strong initial base to market for SageMaker. Cloud service competitor Google has spearheaded developing AI; however, its Google Cloud Platform lags behind AWS in terms of enterprise users, which may play a role in limiting uptake of Google ML as a service product.

The success of the platform could be even more frustrating for the major supporters of the ML frameworks. Google and Facebook have poured significant resources into accelerating the pace of development of the TensorFlow and Caffe2 frameworks, respectively, and making them open source. If Amazon comes in and steals a chunk of the market that directly leverages these efforts, it will be frustrating, especially as ML service platforms are at the early forefront of enabling entities to monetize their knowledge of ML frameworks outside of their internal applications.

In truth, this is a difficult market to forecast, even for ABI Research’s visionaries. SageMaker was only released in December 2017; Algorithmia only launched its AI Layer platform a month prior; and Microsoft’s Azure Machine Learning Studio saw a significant update and refresh last September. Ultimately, the usual suspects of cost, brand perception, usability, capability, and user support will play a determining role in the market share these ML service platforms stand to gain. The development of ML service platforms will significantly impact the speed at which companies develop and implement ML algorithms in their operations. They provide the link between the developments achieved in AI frameworks to systems integrators and developers, making this an area of critical interest for anyone interested in the commercial adoption of AI.