17 Jan Google, Microsoft, And Amazon Place Bets On AI In The Enterprise

Google just announced significant enhancements to its machine learning services (MLaaS), attempting to close the significant competitive gap that Microsoft  has enjoyed, in my opinion, for the last year or so. Not to be left out, Amazon.com AWS announced the company’s own new MLaaS tools and services at AWS Re:Invent last November, trying to court AI application developers to build their smart apps on the AWS cloud. MLaaS is still in its infancy today, but it may become a dominant AI platform for enterprises who would prefer to leave all the messy details to someone else, and rent AI services by the click. This article summarizes each company’s strategies and tactics and tries to size up the winners and losers.

MLaaS: the promise and the problem

Machine Learning is just plain hard, especially the complex area of Deep Learning. Deep neural networks are trained with millions of data samples,  analyzed by racks and racks of NVIDIA GPUs, in order to extract and recognize features and categories. This is the dawning of the “AI Era,” so enterprises and government agencies are of course scrambling to figure out what they need to do to avoid missing out on The Next Big Thing. To get there (wherever “there” is), they must decide which projects to fund, hire rare talent, buy a ton of servers and GPUs, preen their data for the task of supervised learning, and then build and optimize their own Deep Neural Networks (DNNs). Sounds hard? Well, MLaaS provides an easier option: take a short-cut and use pre-trained neural networks for image, video, voice, and natural language processing, offered by the major cloud service providers. Why spend all that time and money training a neural network yourself, when you can just write a cloud-based application that accesses a pre-trained network via a simple API?

Google, Microsoft, and AWS: different strengths and approaches

Google Cloud AutoML provides dashboards to enable the developer to easily gauge the precision of the AI model.

Google MLaaS 

    • Strategy: Leverage Google ’s leadership expertise in AI and Deep Learning (the company has over 7000 AI projects underway in-house, and over one million AI users globally) to provide the most advanced development tools and highest performance hardware platforms for AI development. It is all about the developers since Google doesn’t own the users like Microsoft .
    • Tactics:
      • Make TensorFlow the king of AI hardware and software.
      • Apply AI to the development of AI. Google claims its recently announced Google Cloud AutoML can greatly simplify the complex tasks of DNN development. Instead of augmenting a pre-trained API with additional custom data (as Microsoft offers), Cloud AutoML builds a custom Deep Learning model, starting with the customer’s own data. AutoML comes with really cool dashboards so you can easily see the efficacy of the model as you develop and tune it. Google even provides in-house data tagging as a service—a manual process which some people believe will eventually be automated by AIs.
      • Broaden Google ’s reach beyond the data center into edge and consumer devices and autonomous vehicles. Capture the entire spectrum of AI development on the Google Cloud Platform.
 Microsoft MLaaS
  • Strategy: Use Microsoft ’s massive enterprise and government installed base and its extensive portfolio of productivity and business process tools to become the default provider of ML technologies in the enterprise.
  • Tactics:
    • Provide a wealth of Machine Learning APIs to process every data type, since each company or agency’s data is distinct to their business. Enable the user to extend the trained neural network with data samples that encompass the organization’s products, people, vocabulary, etc. ( Microsoft was the first company to go down this path, and now offers 29 APIs—many of which support customization of the DNN training data).
    • Provide the highest performing Machine Learning Framework for those customers who need to build their own Deep Neural Networks, especially for natural language processing.
    • Enhance every Microsoft product with AI—provide smart features to Office 365, Dynamics, Windows, and eventually every product in the Redmond vault.

Amazon AWS MLaaS

    • Strategy: Use AWS’s extreme scale and a rich toolset to provide the most cost-effective development and deployment platform for AI applications.
    • Tactics:
      • Start by providing the tools and platforms that were developed for Amazon’s massive online business as services on AWS. Tools developed for Alexa and for Amazon’s own eCommerce are now available to help you easily build a chat-bot or voice-activated product or service.
      • Provide world-class development tools such as the MXNet framework, Lex, Rekognition, and SageMaker to ease the development burden. These tools are all very sticky, ensuring that AWS will be the deployment platform after the development process is finished. SageMaker is especially interesting, offering a fully-managed platform for the entire machine learning development lifecycle.
      • Provide the most cost-effective cloud infrastructure for every developer, regardless of which CPU, GPU, or AI Framework the developer selects.
Conclusions
First, I must provide a caveat: however good these AI services may be, enterprises need to be cognizant of the limitations of MLaaS. The problem, of course, lies in the details. What if the pre-trained network as a service does not adequately encompass the universe of faces, vocabulary, and objects you want it to recognize? What if you want to run the AI application on your own infrastructure, keeping all that valuable data in-house where it at least appears to be safe? In either case, MLaaS may not be the AI on-ramp enterprises are looking for. Microsoft and Google are attempting to address these functional limitations of MLaaS, but I think that Google ’s approach could produce more accurate results—AutoML is actually building a custom AI model, instead of simply providing a customizable pre-processing layer.

Incidentally, I was surprised to learn that AutoML runs on NVIDIA GPUs, not the much-vaunted Google TPU (yet to be made available on GCP). I expect that could change soon, but it does indicate that Google ’s pre-announcement of Cloud TPU last spring was timed more to disrupt NVIDIA GTC than to align with impending availability.

Nonetheless, I believe that Google’s bench strength in AI will help the company meet and possibly exceed Microsoft’s current leadership in MLaaS technology, while Microsoft ’s strength in the enterprise software market will help it monetize its AI investments through its applications portfolio.