I have a PhD in AI and I'm having trouble figuring out exactly what's being announced here. Like, check this out:
> Building on advanced research in program synthesis (PROSE) and data cleaning, we have created a data wrangling experience (Figure 1) that drastically reduces the time that data scientists have to spend in transforming data for machine learning.
Huh?
> Figure 1: AI-powered data wrangling in the workbench learns from examples and automatically synthesizes code for data transformations using program synthesis technology.
What?
> Models can be containerized in Docker and deployed to network edge devices, allowing models to score closer to the event and in real-time. Local docker deployments can be used for debugging, while for scaled out production serving of AI, these containers can be managed with Kubernetes, using Azure Container Services.
I've seen this demo at a meetup and its actually quite cool. You use a graphical interface to connect boxes of data transformations, play with the settings, and it will generate a python script that will perform the exact operations you specify to the data in question.
What you're describing are tools for data engineering - sanitizing real-world datasets to be fed into models. Researchers do not have to deal with this task as much because they work with well-defined datasets to provide fair comparisons of the algorithms they develop. Industry is the opposite - the algorithms are usually formulaic and well-defined, but the data itself is not.
Honestly 'data wrangling' is indeed an unfortunately large part of ML work, so I'm interested. I have wondered how tractable making a data transformer that simply shaped and filtered the data based on example and a few constraints would be. Remains to be seen how good a job this has done.
Getting data into a reasonable format is (unfortunately!) a huge part of the machine learning process; that was the motivation for tools like pandas (in Python) and dplyr (for R). Joseph is describing a machine-learning-enabled way to automate some of that data cleaning, which is pretty cool.
I'm reasonably certain that a lot of this is sales jargon. They'll point enterprise CIOs with degrees in music to this site and they'll see a lot of buzzwords and buy.
Currently reading the work of Hubert Dreyfus. Dreyfus was widely ridiculed in the 60’s and the 80’s for coming out against AI simply on the grounds that it wouldn’t work, but for those decades at least he was certainly right as it turned out. Reading his book “The Power of Human Intuition and Expertise in the Era of the Computer” from 1983 is haunting and eerily sounds a lot like today.
>“AI entrepreneurs and researchers will climb a tree, sometimes even a tall one, and then tell you they’ve got all the workings of a space program”
That is a weird position to hold. Surely he must realize that it will work some day; or does he believe there is something more to human-like intelligence, that can never be achieved in silicon? I'm not saying we are close, or that the currently available tools are enough, but at some point in the future we will be there.
> “AI entrepreneurs and researchers will climb a tree, sometimes even a tall one, and then tell you they’ve got all the workings of a space program”
Sounds more like a cranky Luddite than anything else. "on the grounds that it wouldn’t work", lovely argument right there. The list of inventions that "wouldn't work" is vast, including trains, aircraft, personal computers, etc
All modern inventions started as feeble/underpowered/clunky inventions
Compare Atari graphics with PS4 ones
Compare the first works with Mnist digit recognition to what Snapchat filters or AlphaGo do
But that's fine since skeptics means less competition to those who actually make things work
Current-generation ML techniques have proven business value. You can talk to your phone, and it understands a decent set of common actions to be performed. It is now mass-market.
Businesses see the value of the existing ML systems, and can also see that investments in incremental improvements will pay dividends. And so, investment will continue.
Meanwhile, we're finally starting to figure out the actual architecture that allows us to think and move. [1] With continued research, we'll see progress there as well.
Sounds like Microsoft is trying to make a big push specially considering the 8000 strong number in their AI division. Interesting juncture as they seem to have an opportunity to catch on. I hear there is a holy war within Microsoft where one sanct feels Windows and Office are still the bread and butter and therefore all AI related investment should go in the direction whereas the other sanct is pushing towards independent offerings through Azure like Cognitive Services,
I interned at MSFT working on Word. Engineers and PMs in Office see a lot of opportunity in AI. PowerPoint Designer and the newer ML based proofing in Word are good examples of things that have already shipped. I never saw anything resembling a war.
This is awesome! I've been waiting for someone to release an Excel-type ML product to make machine learning more accessible. This looks right up that alley, and will probably "democratize" access to ML in a number of fields that tend to be less coding-savvy.
Something about me feels that these efforts by IBM and Microsoft around AI are less around providing Open Source tools to democratize AI and more around providing "big data" style tools to "big enterprise". Both companies made TONS of $$$ selling Business Intelligence tools (SQL Server Analysis Services, Cognos, etc). They are smart enough to see the danger in open source tools like TensorFlow, Spark, etc. cutting into their lucrative revenue streams in the enterprise.
In particular, Microsoft has always been great about providing tools for no to low cost at the entry level, to get you (or more likely your company) hooked into the ecosystem. Not making a criticism, they have made some great stuff over the years (see the Visual Studio ecosystem for example).
The other angle is providing these tools, which can be complex to install/configure/manage, as a service offering via a subscription as part of the Azure platform. Recently MS has been hiring every superstar/rockstar evangelist/advocate/architect/engineer/etc to help design/build/promote/advocate for Azure that they can find (See Jessie Frazelle, @catie, and a ton of key people in the Golang world). Microsoft isn't just coming to play, they're playing to win.
IBM and Microsoft can embrace the open source tools Tensorflow and Spark by offering Enterprise Support (I know IBM has made a large investment in Spark). Their competitors would be databricks and Google. By being simply not Google that could win over people. Also, both of them need something to differentiate their cloud offering from Amazon.
Access to AI/ML is already very democratic. Takes little time to set run Tensorflow from a docker image, or learn from Jupyter notebooks, etc... with fully open source projects where you can consult the source code and see how an algorithm is implemented.
In contrast this involves running a proprietary operating system, IDE, closed source, etc... Quite the contrary to anything that could be considered democratic.
Then, ML is all about volume. Open a spreadsheet with more than 10000 rows in Excel and see it squirm in pain.
Takes little time to set run Tensorflow from a docker image, or learn from Jupyter notebooks, etc... with fully open source projects where you can consult the source code and see how an algorithm is implemented.
This is exactly the problem -- what you're describing is not easy for anyone outside of tech to do. If you want to, say, run a simple text classification task and have thousands of labels, this is way overkill. Machine learning has the opportunity to become a common place utility for automating repetitive tasks, and the barrier to entry does not need to be learning Tensorflow, Docker, and Jupyter.
Their Text Analytics offering is still for some reason behind IBM Watson (and I don't even like Watson). Missing: named entity recognition, and multi-label classification (if they did a hierarchial multi-label classifier, the would be amazing).
MS is heavily investing on AI because they missed the opportunity of being #1 on the web and on mobile.
However if you go back some decades you will see that Microsoft did have smartphones, before the iPhone, just they weren't as appealing as a product.
This time around I predict it will be the same. When you look at the product (e.g: Visual Studio tools for AI) it looks very featured but not very organized... Microsoft needs to understand that more and more features doesn't mean more perceived value.
Note that the integration is in VS Code, the lightweight Editor++ which is cross-platform. In general most of the AI/ML related work has been & will likely be cross-plat.
Many people doing AI/ML right now are not using Windows. By asking them to move to Visual Studio you are implicitly asking them to move to Windows. Only by doing that you are making it extremely difficult to adopt such technology.
> Building on advanced research in program synthesis (PROSE) and data cleaning, we have created a data wrangling experience (Figure 1) that drastically reduces the time that data scientists have to spend in transforming data for machine learning.
Huh?
> Figure 1: AI-powered data wrangling in the workbench learns from examples and automatically synthesizes code for data transformations using program synthesis technology.
What?
> Models can be containerized in Docker and deployed to network edge devices, allowing models to score closer to the event and in real-time. Local docker deployments can be used for debugging, while for scaled out production serving of AI, these containers can be managed with Kubernetes, using Azure Container Services.
English, Microsoft. Do you speak it?