Microsoft Azure AI — the right platform for your machine learning ambition?
Key aspects you should consider to evaluate the Azure platform for your machine learning applications
In this blog post we will analyze the most important questions about Microsoft’s AI offering, which you as user and decision maker should ask yourself, if you consider to realize AI applications with their product portfolio.
- What are Microsoft’s key products and services for machine learning?
- Do the products and services offered fit to my own AI strategy?
In further articles we also deconstructed
What are the key products and services for machine learning?
First, let’s look at what Microsoft has to offer in terms of machine learning products and services. As this is a field that is continuously updated and revised by Microsoft itself, here is just a brief introduction about the core functionality and purpose. Product names might change over time but as of beginning of 2020 this is state of the art:
Cognitive Services
- Pre-built services and easy-to-use modules to infuse AI capabilities, e.g. for translation, image recognition, transcription via RESTful APIs
- These pre-trained algorithms allow very fast access to ML capabilities
- Customizability and in-depth functionality for highly specified or unique areas of application are rather limited
- It is meant to be the introduction “playground” to discover and learn about machine learning
Azure Machine Learning
- The core data science cloud service to build, train and deploy machine learning models
- Offers convenient visual interfaces to combine open-source or pre-built models and transformation tasks via drag-and-drop into a data pipeline
- In-depth model configuration and tuning as well as off-the-scratch coding in various programming languages like R or Python are also possible
Azure Bot Service
- Managed service to build and deploy conversational agents for a variety of application areas
- Open-source SDK derived from the development of Cortana to build individualized Q&A bots, virtual assistants and the like
- Access natural language capabilities of the Azure cloud and deploy the service throughout different communication channels and messengers
Azure Cognitive Search
- Focuses on knowledge mining from unstructured data stored in files or on the web
- Uses ML techniques to index and filter information from videos, images, pdf-files and the like
- Key functions include OCR, translation, key phrase extraction, location, fuzzy search-term detection, which can be applied to various file formats
Azure Databricks
- Part of MS’s cloud data warehouse ecosystem
- Databricks on Apache Spark allow setup, preparation and training of large amounts of data
- Essential element for using near-/real-time data or streaming high-scale IoT data
Data Science Virtual Machines
- Cloud-based solution to set up a data science working station
- Does not require fully integrated data warehouse landscape
- DSVMs can be used for all kinds of applications for SMEs, experiments or as flexible starting point for more advanced users
The overview shows that the product portfolio covers the entire range of technical requirements for the implementation of machine learning in business or enterprise applications. At the same time, there are smooth transitions between very simply structured entry-level services and advanced requirements.
An intensively discussed topic in this context is AutoML, meaning automation of the applying machine learning to real-world problems in an overall “managed service” from the raw data set to the deployment and maintenance of a machine learning model.
The ambition is to enable also non-technical end users to apply ML to everyday problems and thus increase productivity and output quality. In perspective, this can certainly be of great importance for Microsoft, which enjoys extremely high product and service penetration with direct end-user touchpoints like hardly any other competitor.
However, these are often not tech or machine learning experts, but simple productivity tool or operating system users. Only Google has a similar user profile, although not as strong in the professional / enterprise area as Microsoft.
Do the products and services offered fit into my own AI strategy?
Microsoft’s offering is quite comprehensive and ranges from entry-level to professional data science tools. The core product is Azure Machine Learning, including the ML Studio, which is intended to provide a complete service suite for data scientists and machine learning engineers. As always, it is difficult to give a general answer to the question, whether a solution is a good fit or not. It strongly depends on what type of user you are and what exactly you are looking for. But let’s dive into the rather generic characteristics of the assessment, before tackling the issue of an individual perspective.
#1 Industry focus
With regard to the application focus, the Microsoft AI platform is rather generalistic and does not outline highly specific industry solutions. In terms of communication, six high-level verticals of particular importance are mentioned:
- Government
- Financial Services
- Retail
- Manufacturing
- Health and life sciences
- Energy
These are primarily showcases of large customers who have implemented specific use cases built on the Microsoft Azure Platform. However, no ready-made industry solutions are available on a large scale outside the Azure Marketplace. In almost all products and services, Microsoft tends to be the generalist which is trying to serve the broadest possible market with its solutions. That does not mean that industry-specific solutions are out-of-scope for Microsoft, but they tend to outsource “extreme” customization to their broad network of implementation partners and focus more on the infrastructure that is required to run such services.
#2 General strengths
Machine learning tools:
- Visual user interface of the Machine Learning Studio enables smooth usability even without very deep technical know-how
- Comprehensive support of common frameworks and development tools
- Integration possibilities in Office, PowerBI, Teams, OneDrive, Skype, LinkedIn etc.
- Microsoft AI School and comprehensive content offering (including video series) on AI and data strategy and cultural transformation of the company
General infrastructure:
- International coverage and support levels, possibility to test many services free of charge as well as a broad network of implementation partners
- Clear and well-structured service offering with comprehensive documentation and support
- Azure Marketplace as tool platform and sales interface (currently still comparatively insignificant)
- Broad database support (relational, distributed systems/data lake, data streaming)
- Azure Data Factory as a modular system for completely cloud-based data warehousing and data pipelining
- For enterprise-related analytics applications, PowerBI provides a strong front-end tool compared to Google’s and AWS’s offerings
#3 General weaknesses
Machine learning tools:
- Flexibility and usefulness of the pre-packaged Cognitive Services — have more of a showcase character than actual user benefit
- Comparably low support of complex deep learning models in Azure ML — primary focus of the managed service is on “simple” statistical analysis and machine learning techniques (regression, decision trees, random forest)
- Individualisation options and range of functions through “drag & drop” interfaces sometimes less flexible than with competitors such as AWS
General infrastructure:
- Expandable data governance, data quality and master data management (not available at all) solutions in portfolio
- Little or no specialization in specific verticals / industry-specific solutions
#4 Product-problem-fit
As already mentioned, a lot depends on the expectations and objectives with which the search for a specific solution begins. General statements are therefore difficult to make. Let us now look at the above described offering from the perspective of four different user groups and draw a concrete conclusion about the suitability of the tools and services for each of these groups:
- Productivity users (entry-level or no Analytics/AI knowledge at all)
- Data analyst (intermediate Analytics/AI knowledge)
- Data scientist (in-depth Analytics/AI knowledge)
- Machine learning researcher (top-notch Analytics/AI knowledge)
Productivity users
No specific technical understanding around AI, software development or statistical analysis. End user of everyday productivity and communication tools, who should be enabled to optimize his personal working day after a short training by “hacks”. Simple and existing forms of these include Excel macros, RPA applications or connector functions within the Office producitivy suite or through tools such as Zapier.
Fit: low
Reasoning: AI is still in a comparatively early stage of development, many off-the-shelf applications for the mass market do not yet exist, or if they do, they are so conveniently automated that the end user does not even notice that AI is at work here. Cognitive Services is best suited as a drag & drop solution for even the most untrained users to conduct their first experiments. However, it will not be possible to derive truly productive solutions from this. The pre-trained standard models are too generic and imprecise. Furthermore, even the basic understanding of AI or level of data literacy is still very low for the average end user. This may change in the next 2–3 years, but until then it remains a gimmick for this user segment.
Data analyst
Experienced handling of data, databases and common analytics tasks. Ability to carry out simple to complex analyses by themselves and prepare the results using suitable reporting or visualization tools. Strong focus on putting business data into the right context and interpreting it in order to derive recommendations for specific business decisions.
Fit: Very high
Reasoning: Data or business analysts in the broadest sense are enabled by the comparatively simple user interface of the ML Studio to carry out analyses and experiments with data, which otherwise require higher entry barriers through corresponding software or in-depth tool knowledge. Due to the easy integration into standard tools like Excel and PowerBI, analysts are able to integrate machine learning methods into their daily analysis routine. The focus here, however, is on the processing of medium-sized data sets with methods such as multidimensional regression analysis, random forests or support vector machines. Setting up neural networks and deep learning models will currently not become the core application area for this user group.
Data scientist
Data scientists focus on the explorative analysis of data sets using advanced statistical methods. For data scientists, both the ML Studio and infrastructure products such as DSVMs or Databricks provide useful support as necessary end-to-end process modules. The focus of expertise is clearly on the generation of relevant and non-obvious insights from large and complex data sets. Accordingly, data scientists are dependent on being able to apply their diverse knowledge of data analysis as efficiently as possible through structured experiments.
Fit: High
Reasoning: The greatest advantage for data scientists is to work with deep methodological knowledge in an easy-to-use and efficient user interface. This user group benefits more from the customizability of the models. Another advantage is the integration of data engineering tasks (preparatory measures) as well as the possibility for immediate deployment of models, without time-consuming handover processes to software developers.
These tasks ideally should not be mentioned at all as part of the job for a data scientist, but in reality data scientists turn out to be full-stack engineers and developers trying to bring their passionately developed solution to production due to the lack of supportive resources. For newcomers and lateral entrants to the field of data science, the Microsoft toolbox certainly offers many advantages and comprehensive documentation — but it is not yet entirely code-free work with data sets.
Machine Learning Researcher
This role characterizes the advanced data scientist with a strong focus on machine learning up to machine learning architects and engineers who want to develop high-end models for dedicated use cases and use them productively. Not surprisingly, the choice of methods often ends up to be neural networks — especially in the area of processing unstructured data. The models must be able to process extremely large amounts of data and productive use, for example in the context of safety-relevant applications, requires extremely reliable outputs and monitoring of performance metrics.
Fit: moderate
Reasoning: Azure AI naturally claims to be equally suitable for all users. But just as graphic designers do not use PowerPoint, modern database systems are not based on Access and company-wide real-time reporting should not be displayed in Excel, the AI platform falls somewhat short of the demands of the absolute high-end users.
Deep learning models are often very expensive in setup and training, especially when large amounts of multimedia data are used. Furthermore, they are not built using simple drag and drop logic, but require conscientious detailed work and continuous testing and tuning. Here the central advantages of the simplistic user interface just don’t play out. Ultimately, this assessment is of course subjective, but the predominant impression is that the Azure AI platform as it is today — and in line with the formulated corporate strategy “bring AI to everyone” — is not geared towards the high-end user, but more towards easier access to the topic of data science and machine learning for a broader user audience.
Conclusion
Azure ML scores above all with a strong front-end, easy access for users who do not want to get used to the full technical diversity of an AWS platform and well-structured, extensive documentation and learning materials.
However, throughout the entire service portfolio, the application focus is on rather simple and standardizable use cases that apply to a wider range of companies (recommender systems, personalization, simple object recognition, translation, knowledge extraction and chatbots).
Cutting-edge deep learning models for industrial applications or highly specialized use cases can also be mapped technically, but hardly benefit from the advantages of the system mentioned above, which clearly aims to reach “everyone” first and accordingly focuses on solutions for everyday problems on a smaller scale.
As far as pricing is concerned, it’s almost impossible to make any statement at this point, as this is extremely dependent on the chosen configuration in the case of the usual pay-as-you-go models, on the one hand, and is updated very regularly by the provider on the other. The Azure cloud platform generally offers very granular selection options and billing cycles. However, in the case of larger projects, much also depends on individual negotiation or discounting — for example, if services are used over a longer period of time, the percentage of discount can quickly be in the mid double digits.
If you want to kick-off a test project, I’d recommend to get a trial version (first couple of days are free of charge) and build up your own mind. Furthermore, check some detailed reviews from fellow bloggers to get an initial “hands-on” impression:
