Frameworks are only an intermediary step to the wider adoption of machine learning in applications. What's needed are more visual products and those are still a couple of years away.
The current machine learning (ML) focus on frameworks is a middle step in the needed evolution of the productization of ML and its inclusion through the application environment. In order to truly succeed, the ML vendors need to think more like a business user and less like a programmer. One way to start is to learn the lesson the business intelligence (BI) sector provides.
There is an aphorism that history doesn't repeat itself but it rhymes (often attributed to Mark Twain, but there's no proof he said it). When it comes to the adoption curve for machine learning in business, it has a ring of truth. Deep learning (DL) frameworks, such as TensorFlow and Caffe, are getting a lot of technical press coverage, because that's exactly what they are - technical. To understand their limitations, let's review the standard definitions of computer language generations:
- 1st: Machine code, binary 0s and 1s.
- 2nd: Assembly language.
- 3rd: Logic coded in text, what most of us consider to be a computer "language," such as Python, C, Fortran, and Cobol.
- 4th: Environments with modern user experiences (UX), such as Business Intelligence (BI), Visual C, PowerBuilder, Oracle ORCL -1.62% and SAP and other development tools.
Each generation allows more programming to be done (as measured by debugged lines of code) with less work. Most importantly, each level allows a wider group of people to accomplish tasks, because each level is a layer of abstraction that hides the gory details of the layer(s) below. Much of 4th generation work is focused on allowing people to "program" without coding.
Another term often used is that of the magical "data scientist." The main problem with the myth is that such a person should exist in the long term. I believe the phrase should refer to teams of people in the early days of a new technology solution, when in-depth knowledge is required to solve problems.
However, take the first term in the 4th generation list: BI. Originally, adding analytics to existing business applications required in-depth knowledge of statistical modeling in order to code that information in a 3rd generation language. This is when people first began referring to data scientists.
As BI advanced, graphical development and insight delivery enabled a more accessible UX that widened BI's appeal and user base. Now the question for BI is "how far up the information and insight delivery chain does the ability to generate analytics go?" BI applications let business analysts create wonderfully detailed analytics and are allowing business management to do more by dragging a few objects on a screen. Graphical UX enabled the explosion of BI in business applications.
In historic rhyming terms, frameworks are 3rd generation tools. They've hidden some of the details of machine training, the equivalent of 2nd generation machine knowledge, and abstracted that in modules and hidden code, allowing programmers to extend their applications with some ML techniques and abilities.
Cloud, Not Frameworks, Driving Current Machine Learning Acceptance
What's allowing ML to have the major impact we've seen is not the power of frameworks, it's the power of the cloud. The move to the cloud has allowed core programs to have greater reach than ever before. Amazon Alexa and Echo, for instance, rely on Amazon AWS to drive the applications. The ML applications being talked about the most are very few, but they're very prevalent - thanks to cloud servers.
However, most businesses do not yet run on the public cloud. That will eventually happen, but it is still a few years away. There's too much software running on servers inside the firewall. The first step most companies will take to the public cloud for their mission critical applications will be to deploy private cloud (cloud infrastructure and server farms that remain inside a corporation's firewall). The result will be more controlled and more customized systems than found in the public cloud.
Meanwhile, the addition of ML to applications will be done by the few highly-paid experts who understand ML frameworks. Therefore, enterprise ML will focus on cloud applications rather than hiring expensive, internal resources.
Business Intelligence Growth Should Inform Machine Learning Growth
Last decade, 4th generation BI tools began to gain traction, replacing earlier generation tools. IBM IBM -0.65% bought Cognos, SAP bought Business Objects and Tableau began a strong effort to change their software to become more visual, both for delivered analytics and in the development cycle.
The changes driven by younger BI firms' embrace of both UX and the cloud has transformed the face of BI in the last decade. Visualizations are everywhere, and management can do far more to investigate their own businesses in real-time than they could before.
The same change is needed in ML. Containers are barely a start, enveloping ML code and environment in what you could think of as "modules on steroids" but still requiring 3rd generation knowledge to leverage their power.
What I'm looking for is for the rise of the first companies who begin to provide visual tools to add ML into the existing visual development environments. That will allow vastly more people to take advantage of ML without having to leave their own areas of expertise and become machine learning experts.
Given how many frameworks I see and not a mention of even an early 4th generation tool, I expect that the development of development environments.
I've heard too many analysts in the BI space still, to this day, talking about how the BI challenge is to teach the business user to think like the data scientist. They argue about ways to train the business user to think more like a programmer. That is backwards.
The challenge is for the data scientist to think like the business user. The same is true for the ML experts. The faster they begin to think like the business user, and to develop tools that will let business analysts and management leverage the power of ML without having to think like an "ML scientist", the faster the potential of ML will become reality.
-- The author and members of the TIRIAS Research staff do not hold equity positions in any of the companies mentioned. TIRIAS Research tracks and consults for companies throughout the electronics ecosystem from semiconductors to systems and sensors to the cloud.