Voice of the DBA

Tag: machine learning

  • Smarter Design

    I have build a lot of PowerPoint decks over the years. Many of presentations, some for explaining a process, and once in awhile, I use it to create an image that’s simple, but beyond my skills. Even putting some geometric shapes together in a clean, aligned manner, is something I’ve found easier to do in PowerPoint than Paint.Net.

    Some of you may have noticed the PowerPoint Design Ideas that pop up when you create a new slide. At first I didn’t pay much attention, but I’ve started to take advantage of the ideas at times. The way that slides are put together is often much more visually appealing than any I’d think of by myself. Sometimes I take an idea and modify it slightly. At first I thought most of these were just standard templates, but Microsoft is starting to use more ML in the examples that pop up on the side.

    There’s a good AI Show episode at the bottom of this post that explains some of what they do. From image recognition to intelligent cropping to natural language processing, they are finding ways to make better design ideas available. the episode is only minutes and worth watching if you want to know more about how they use technology in their recommendations.

    To me, this is a good example of how technology and humans work together. Someone needs to come up with some good designs, based on certain situations. Users then provide them feedback on if they’re useful or not, and they can modify things. I even learned how to change the icons for the designs if I don’t quite like them. My changes (and yours) get fed back to new models as training for the future.

    Would this make all presentations look alike? I think there is some danger here that we start to coalesce into certain patterns. We might see lots of slides with dates become timelines, but they won’t be exactly the same, and there are different ideas based on content. Even the words in the titles. I made two slides, same content with a different title and I got slightly different recommendations.

    I think AI and ML have good futures, but not as the sole decision makers for how to react to data. Instead, I think humans constantly providing feedback, input into models, changing weights, and yes, helping with more complex things like design, will improve the results that more of us get from our tools.

    Steve Jones

    Listen to the podcast at Libsyn, Stitcher or iTunes.

  • Giving Computers Ethics

    I was reading a fascinating paper recently about autonomous cars. I’m actually excited about having a car that can drive itself, though I think this is likely quite a few years away, despite the hype. Ever since I read Red Thunder, I’ve thought that we would first get full time autonomous cars that would either be limited in where they were in use, or part-time autonomous cars that could only be self-driving in certain places. Dense inner cities, or maybe isolated highways might be good places to try this, in my mind.

    While we want to do some programming of these cars, we also have a lot of AI/ML systems in place that run models trained to react in certain ways. They identify things that are moving and stationary, trying to determine how the car should navigate and react. The systems aren’t quite as tightly programmed as many of us expect, with if this then that logic. Instead they have guidelines that are decided upon by the designers and then reactions to data inputs and analysis are a little more fuzzy.

    What are the goals? Well, in most cases they are just moving the car safely down a road. In crisis situations, it’s a little more murky. What happens when collisions are unavoidable? How should the car react? Humans often panic and do strange things, but we don’t want erratic behavior from automated systems, so what should we set as goals? There’s a bit of research that was done to ask humans what they would do when they can consider the situation a little more slowly.

    In short, humans make different decisions in different cultures. There are clusters and tendencies in different parts of the world, which is interesting. While people are people and behave similarly in many cases, we tend to value different things, depending on our views of the world. That can be problematic when we start to expect computer systems to be more consistent or predictable. After all, we should decide how computers react and be able to trust our decisions are followed. It is up to humans to imprint our ethical desires as a society on computer systems.

    This is an area where I feel AI and ML systems are moving faster than our ability to comprehend the implications. I would want to have a framework built for automated systems, certainly cars, and then expect all vendors of systems would implement that framework in their vehicles. However, this goes beyond cars, and in any places where we are using software, AI/ML based or not, we ought to publish a comprehensive outline of the way in which our system works.

    Computers have the capability to improve our world and reduce chaos, but only if we agree on the way in which these systems work, and disclose in a transparent way what data they handle and what decisions they make based on that data. I hope that we start to get better about informing the world the goals and operation of our systems. I’m not sure that will happen anytime soon.

    Steve Jones

    The Voice of the DBA Podcast

    Listen to the MP3 Audio ( 5.6MB) podcast or subscribe to the feed at iTunes and Libsyn.

  • Badly Trained AI

    Most of us know that data is being used to make more and more decisions inside of all kinds of organizations from retail giants to banks to sports teams. We are constantly asked, or see reports, of data driven decisions. We often need to show some data that supports and explains the rationale for making some choice. As our populace becomes more data savvy, I expect this trend to continue.

    AI (Artificial Intelligence), and the related Machine Learning (ML), are becoming more and more widely used. From mobile phones to autos to trading systems, we regularly see new “AI capabilities” being added to products and services. No business or industry seems immune, and I’m sure many of you are seeing AI being incorporated or feeling pressure to start using some AI in your work. As you work with AI, or start to, you’ll quickly realize the importance of data in your efforts.

    This is true for the cleanliness of data, but perhaps even more important in the tagging of data sets. As Amazon learned, building an AI or ML system, is hard. They scrapped one system that was being used to rate resumes and help their recruiters sort through the volume of applications they received. Why? Because of bias.

    Apparently the system would downgrade women’s resumes for various reasons. To me, this is a perfect example of a principle I’ve had throughout my career: garbage in garbage out. In this case it’s not necessarily bad data that was the problem, but bad tagging of what was a good and bad resume, probably from the internal prejudices of a few people.

    There will be more dangers as we use ML and AI technologies in our work. It won’t be enough that we clean the raw data for training, but also that we clean and properly manage the tagging of what data sets represent the results we are looking for. Like in much of our software, it’s easy for us to only consider the happy path, to only tag those items we think are good results. That is useful, but we might also be unconsciously tagging other results as bad, which appears to have happened to Amazon.

    We can build systems that do a better, more rational job than most humans, but we need extraordinary care to ensure our training data lacks bias. Unfortunately, most people both think they’re not biased and are unwilling to spend extra resources to deeply examine the data. Two things that worry me about the future of our AI/ML systems that will inform us.

    Steve Jones

  • Machine Learning in the Database

    When SQL Server added the ability to execute R code, the decision seemed to split the customer base into two groups. One group was impressed and thought the idea of executing R code to analyze data in the database was a good idea. They were excited and impressed by the loan classification demo. If you haven’t read about this or seen the demo, it’s very interesting, and it’s something you might take a few minutes to read or watch it.

    The other group of customers felt this was a poor use of CPU cycles for a very expensive SQL Server CPU license. Running a complex analysis, training models, and other functions commonly associated with R scripts aren’t a good use of scarce resources. They would rather have R code execute on a separate server, much like any large messaging workload might be better served by a service such as AWS’ Simple Queue Service rather than Service Broker.

    I tend to be in the first group, as is Dr. Low. He writes that there is a place where Machine Learning Services (MLS), with both R and Python, are a good use of resources. Not in all cases, and certainly not for all work. The difficult parts of training models and doing the hard work of coming up with new ways to perform an analysis is definitely better left to workstations and data scientists. Those actions might not be worth the resources they take.

    Once the models are trained, however, the executable load of submitting parameters to a model and getting a prediction is small. SQL Server allows us to load pre-trained models into the database and just call them as needed. Plus, the R models run in a multi-threaded fashion, unlike the single threaded execution in clients such as R Studio.

    As with any feature of SQL Server, it’s important to test and evaluate the real world impact of new code on production sized workloads. Not only will you want to measure the load of your model execution, but you should also measure any changes in your existing workload with the additional R or Python code load. While I wouldn’t prevent the use of MLS in SQL Server, just like SQL CLR code, I would be careful about introducing without extensive testing, including dark deployments and simulated loads.

    Steve Jones

    The Voice of the DBA Podcast

    Listen to the MP3 Audio ( 3.2MB) podcast or subscribe to the feed at iTunes and Libsyn.