Evaluate Weigh the pros and cons of technologies, products and projects you are considering.

Comparing MLaaS providers by cost, UX and ease of use

MLaaS allows companies to add machine learning capabilities without software development. There are still some barriers to entry, however, and providers are not one-size-fits-all.

It's not long ago that having a machine learning platform was obviously a market-shifting advantage, but not necessarily an essential. Today, however, machine learning is a must-have for the competitive enterprise. There are potential barriers to implementation that dominate the decision-making process.

The biggest of these are cost and ease of use. Machine learning platforms are often difficult to integrate with in-house systems, and many are hard to master. Machine learning as a service (MLaaS), however, is a technology that offers relief in both areas.

MLaaS is already established as one of IT's fastest-growing markets. Valued at just over $1 billion four years ago, it is expected to top $20 billion four years from now. A large part of this growth is attributed to the parallel surge of the enterprise IoT market, where effective adoption is often dependent upon machine learning resources.

The biggest players in MLaaS providers are well-known -- but this roundup analyzes four key elements that separate platforms: cost, integration, built-in language compatibility and user-friendliness.

IBM Watson Studio / Watson Machine Learning Cloud

IBM's extensive Watson suite includes MLaaS functionality augmented by a wide range of development and management tools. Intended for use by developers and data scientists, it's based on hands-on models created in Watson Studio, and managed via OpenScale. Cloud Pak may be purchased separately to automate AI lifecycle management. Models can retrain dynamically.

Watson Studio's visual modeling tools make it convenient to quickly generate insights. Notebook tools friendly to R, Python and Scala facilitate analytics for data scientists. A neural network modeler and flow editor support developers building machine learning into cloud apps. The Bluemix Dashboard aids developers and data scientists working together on models.

The Watson Machine Learning Cloud works with SPSS and existing algorithms in Spark MLlib out of the box.

Watson Machine Learning is easily integrated into existing systems intended for complex analysis and application development and aimed toward professional users. Businesses-level analysts and citizen data scientists may see issues with UX and user-friendliness.

There are three pricing tiers -- Lite, Standard, Professional -- and the first two are pay-as-you-go:

Lite: Free, 5 models / 5,000 predictions per month / 50 hours training, batch deployment

Standard: $0.50 per 1,000 predictions per month / $0.50 per capacity unit hour

Professional: $1,000 per month / 2,000,000 predictions per month / 1,000 hours

Google Cloud Machine Learning Engine

Based on TensorFlow, the Google Cloud ML Engine capitalizes on the tech giant's considerable SaaS dexterity, with the ML engine extending across a wide range of services. Google AI accommodates natural language processing, translation, image recognition and other growing AI applications, while offering an array of APIs. The Google Cloud ML Engine is integrated with them all.

This MLaaS provider's biggest strength is in deep neural network modeling -- and the tool set is counterintuitively very plug-and-play. An AI Hub enables the creation of AI pipelines and includes an extensive set of out-of-the-box algorithms, a set of building block components (for image/video analysis, language, sentiment analysis), and facilitates smart app development. The AI Platform for machine learning development and implementation includes a JupyterLab-integrated enterprise notebook service for machine learning framework management. The AI Platform also includes preconfigured virtual machines and deep learning containers for rapid application development and can host models as hosted prediction engines.

The AI Platform's integrations include Compute Engine instances that host a range of popular machine learning frameworks besides TensorFlow, including PyTorch and scikit-learn.

The major downside of Google Cloud ML is its cost. Only the AI Hub and the notebooks are free; everything else is by subscription, and many of the fees are negotiated by contract. The confusing complexity of the pricing is partially mitigated by a pricing calculator provided on the Google Cloud website. Though some hourly processing rates are quoted on the website -- for example, machine learning training jobs are billed at $0.49 per hour, per training unit in North America -- you'd have to contact Google for a specific number.

Microsoft Azure Machine Learning Studio

Microsoft moved early to be a leader in the MLaaS market, deploying resources for data scientists, developers and business workers alike. Its ML Studio is designed for flexible and extensible hands-on development and can accommodate the full spectrum of users and applications, from the simplest to the most grandiose.

Microsoft's hands-on philosophy, which has origins in its BI/data warehouse offerings within SQL Server over 15 years ago, requires users at all levels to perform the steps of any machine learning build themselves: Data exploration and cleaning, learning method selection and validation must all be done by the user. This philosophy cultivates competence in nondevelopers and has been successful for Microsoft despite giving its competitors an ease-of-use advantage.

The ML Studio comes with an array of out-of-the-box algorithms, and the Cortana Intelligence Gallery (a community-based library of canned machine learning solutions) is available as a development resource, offering templates for human resource analytics applications, retail customer prediction models, automated support ticket analysis, fraud detection systems, supply chain demand forecasting and hundreds of other offerings.

Like IBM's platform, the Azure studio's biggest downside is the learning curve and required project time. Despite the accessibility of each stage of project execution, less experienced users must make a considerable investment of time and effort to complete a project.

The ML Studio is free to users with a Microsoft account, with a workspace of up to 10 GB storage, including Python and R support and predictive web services. One hundred modules per experiment are allowed, and there is a one-hour limit on experiment duration.

There is a Standard enterprise-grade workspace available to those who have an Azure subscription, available for $9.99 per month / $1 per studio experimentation hour; storage space unlimited, data can be read via on-premises SQL and a production web API is available.

AWS Machine Learning

Like Google, Amazon is a leader in SaaS, which gives its MLaaS providers a dominant market position. Unlike Microsoft's ML Studio, it is exceedingly easy to use, with many wizards and convenient tools that don't require deep knowledge or training to use effectively. Model creation is guided, and the resulting APIs are simple and easy to deploy in applications.

AWS ML offers considerable automation, making it even more attractive to machine learning novices. The service chooses best methods and can even discover categorical columns with no preconfiguration.

Its downside is that no unstructured learning methods are available, so its capacity for deep learning applications is limited.

Pricing is usage-based, depending on the individual service used, making billing somewhat complex. Compute fees are based on a $0.42 per hour rate, with a batch prediction price of $0.10 per thousand predictions.

Dig Deeper on Machine learning platforms

Start the conversation

Send me notifications when other members comment.

Please create a username to comment.