Voice of the DBA

Tag: machine learning

  • Intelligence from Data

    There is an incredible amount of data in the world, and all that data is changing the way industries work. That’s the opening to a keynote talk from Jim McHugh at the O’Reilly Artificial Intelligence conference. The talk is short, 12 minutes, and interesting to listen to as Mr. McHugh looks at autonomous cars and healthcare, talking about the impact of artificial intelligence on advancing these industries. There are examples showing how data and AI systems are already being used to change the way the transportation and medical fields can work.

    Whether you want to see more robot help in our world or not, I suspect some level of this is coming, and it’s being driven by data. We have more and more data, and as companies have success in analyzing this data with various types of AI and machine learning systems, there is pressure for other companies to join the trend and build their own systems. We certainly see that with the push from Microsoft that emphasizes the R Services in SQL Server. At the recent Data Science Summit, there was a demo in the keynote (around 17:00) of over 1 million classification queries per second running inside SQL Server. You can even try this yourself on SQL Server 2016 Developer Edition (for free).

    I’m sure that a few of you will start to get more complex analysis projects inside of your organization. Maybe you’ll help develop some sort of prototype, or maybe you’ll just be responsible for helping get the data to the data scientists. I’m also sure that some of you won’t be thrilled with the results. After all, throwing a bunch of data at a few algorithms and expecting some rapid development isn’t likely to work great.

    At least not the first time.

    One of the thing I’ve seen from many people as I study data science, machine learning, and related topics is that this isn’t a simple process. Building a useful and successful machine learning system requires experimentation, and really, ongoing experimentation, as you examine, clean, discard, and make decisions on your data. In fact, the data preparation might be the most difficult and time consuming part of the process. That’s great, since many of us are the people that will work with the data, but it’s bad in that our management might not want to have the patience to experiment, evaluate, and re-tune their systems, much less wait for data to be well prepared.

    I do have high hopes for many complex problems to be assisted with machine learning and artificial intelligence in the future. I’m glad that companies are experimenting, and I think it’s great that so many data professionals are getting excited by the possibilities. Remember that this field is hard, and requires lots of work. Keep learning and growing your skills, and above all, remember that the scoring against your data is more likely to be closer to a baseball game than a bowling match. A 30% success rate might be amazing and those perfect games are likely very close to impossible.

    Steve Jones

     

  • What is the Future of Monitoring?

    I read Tom LaRock’s post on the future of monitoring, looking at the rise of machine learning and complex algorithms to evaluate systems. Tom notes that the data input into a system is crucial in determining whether the system can produce a prediction or output that has value. Since the humans that decide what data is included may not choose wisely, the system won’t necessarily reduce the time required to manage a system.

    I tend to agree. From my conversations with those people having success with machine learning systems, the data preparation is the most critical (and difficult) step. Shortcut the cleansing and organizing, or make poor choices of what data to include and you won’t likely get actionable results.

    However, I have a different view of monitoring. After my watching of the Darpa Grand Challenge (editorial on this recently), I think the future of monitoring is perhaps actually making changes to code in-line. Perhaps with approval from humans, but I suspect that we can train some process to understand how particular batches can cause issues. We can certainly set filters that might note index changes on large tables could be problematic and should be investigated before an issue arises. We can teach a system to recognize code deployments, perhaps even roll back certain changes if the application fails. We certainly could have a machine learning system watching index usage and query plans to recommend indexes in a real time manner, perhaps even turning on and off on-line rebuilds.

    I think there is possibility, but where I’d really like to see advanced monitoring is not in production environments. I’d like to see better systems that can watch development, helping suggest or rewrite SQL before it’s deployed. Perhaps applications can warning developers of potential performance or security issues. Such systems could help us in building more consistency into our applications. We can have systems that help our developers code better.

    As our databases and software become more critical to the functioning of most of our businesses, we certainly need more reliable and robust development practices.

    Steve Jones

    The Voice of the DBA Podcast

    Listen to the MP3 Audio ( 3.6MB) podcast or subscribe to the feed at iTunes and Libsyn

  • The Black Boxes

    Machine learning and artificial intelligence seem to be the hot topics these days. From bots that can interact with people to systems that learn and grow as they process more data, it seems that science fiction is becoming reality. At least, in limited ways. Autonomous cars, perhaps the highest profile example of these topics, are advancing and being tested in a few locations around the world, but I think we are a long way from having human controlled and autonomous cars interacting freely at any scale. There are still plenty of issues to work out, and the consequences from mistakes require serious consideration.

    I was thinking of this as I read an interesting question: Whose black box do you trust? It’s a look at algorithms and machine learning, and the impact they have on the world around us, despite many of us not understanding how they work. The main examples in the piece are in the area of journalism as it relates to social media (primarily Google and Facebook), but also touches on autonomous vehicles, both autos and planes. The latter was a bit of a shock to me as I assumed humans always handled takeoff and landing, something the author says doesn’t happen at SFO. Some searches around pilot sites seem to note that automated landing is done regularly to test systems, but is used in a minority of cases.

    The question is, do we trust the black boxes that run our systems, and really, does it matter? In the piece, Tim O’Reilly says he has four tests for trusting an algorithm:

    • the outcome is clear
    • success is measurable
    • the goals of the creators align with the goals of the consumer
    • does the algorithm lead creators and users to better long term decisions.

    Those are interesting ways to evaluate a system, though I think the problem is that the last two are a bit nebulous. One of the things that I see more and more as I get older is that the same data or the same facts can lead two different people (or groups) to two different results. Our goals, our interpretation of events, even the weights we place on the various factors in a complex system vary dramatically from person to person. In such a world, can we truly evaluate what the goals of a creator are? Forget about consumers, assume one person building a software system. They will have multiple goals, and do we really think their goals can be easily listed? Or weighted/ranked appropriately? What about when the goals change?

    I really think that the black boxes need more disclosure, though I freely admit there isn’t a good way I know of to do this. However, I do know one thing that can be better disclosed: data. We can have more openness and analysis of data from software systems along with some accountability by creators for the impacts of their software. Again, I don’t know how to enforce accountability, especially at scales that encompass millions of consumers and easily cross country borders. That is a problem I think we need to find ways to tackle, at least at some manageable level. Maybe using the 80/20 rule where 80% of consumers and creators find the outcome to be a good one.

    The world of technology and software are advancing and growing extremely quickly. Certainly hardware advances, but it seems the last 5-10 years have been more about new and different software applications that fundamentally alter the way humans can interact in social, business, and government situations. Underpinning all the changes is data. New data, more data, and novel ways of working with this data in ways that were unheard of 20 years ago. It’s an amazing time to be a data professional.

    Steve Jones

    The Voice of the DBA Podcast

    Listen to the MP3 Audio ( 5.2MB) podcast or subscribe to the feed at iTunes and Libsyn .

  • The Machines are Learning

    One of the technologies that Microsoft is promoting heavily is machine learning. This has rapidly gone from a technology that I heard little about to being in multiple keynotes that I’ve seen at large conferences this year. It almost seems that I can’t go a week without seeing some sort of machine learning article, announcement, or reference.

    How applicable is machine learning for most of us? I’m not sure, but banks have certainly taken advantage of machine learning to reduce their risks associated with fraud and their systems work well. In fact, given the ways in which I travel, I’m amazed that I don’t get more calls about fraud related to my card use. With me in a new state almost every month, on a very irregular schedule, sometimes buying computer supplies far from home, I get called by a bank once or twice a year.

    Across the ten or so calls I’ve gotten in the last few years, only one questioned legitimate purchases. The rest caught fraud on the same day that someone tried to use my card without my knowledge. That’s a very successful rate of both true positives, and a very, very low rate of false positives. I’m impressed.

    If you haven’t played with machine learning, Microsoft has made it easy to give a try in Azure. You don’t even need to put in a credit card to get up and working with machine learning. Whether it’s applicable to your industry or not, you’ll have to decide, but I do think that lots of tedious analysis that humans do now could be better done by a machine.

    At least after one of us humans has configured the algorithms and trained the machine to recognize patterns. And, of course, with one of us technologists periodically tuning the system to work better  and monitoring the analysis as data changes.

    Steve Jones

    The Voice of the DBA Podcast

    Listen to the MP3 Audio ( 2.6MB) podcast or subscribe to the feed at iTunes and LibSyn.