Voice of the DBA

Tag: machine learning

  • A Great Use for AI

    In the last couple of years there has been a tremendous amount of hype for machine learning and artificial intelligence as a way to improve the world. Plenty of companies have tried to implement ML/AI to generate more revenue or improve their products, often with mixed success. However, I recently saw a place where I think AI might shine.

    I’ve never owned a Roomba or any robot vacuum, and I’ve never encountered a poopocalypse scenario. I do have a cat that is an avid hunter, so I certainly could envision something similar with a carcass in the house, but apparently, some owners of these vacuums have had a very messy experience when they pet has an accident and the robot vacuum attempts to clean the floor.

    The company has implemented a camera and AI to try and avoid this happening, as well as avoiding other obstacles. How this will actually work remains to be seen, but it’s a good place to use AI to try and detect objects that might cause issues, notify the owner, and avoid creating a mess when trying to clean one up.

    This is also a place of low impact if the AI doesn’t work perfectly. If the model can’t determine what an object is, avoid it and flag the situation. Allowing owners to provide feedback and improving the model over time is what I’d want to see, with regular improvements that might help the system tell when an object is something that could cause issues. If Roomba does a good job, they’ll use this as an opportunity to gather data and improve their products.

    AI/ML isn’t often a build it and forget it technology for systems. These technologies use models, which are inherently incomplete and don’t always match the real world well. They need a lot of training, with new data, across time to become something that works really well.

    Are they worth the effort for most systems? I don’t know. I do know that good data science is needed, lots of data for training and testing, and a set of boundaries where the system works well and where it doesn’t. I suspect we’ll see more businesses having success with AI over time, but not in all situations. I suspect older extrapolation and human judgment work just as well for lots of problems.

    Knowing when each might work more efficiently will be a challenge for years to come.

    Steve Jones

    Listen to the podcast at Libsyn, Stitcher, Spotify, or iTunes.

  • Where We Need Better AI Disclosure and Responsibility

    There are a lot of contract and gig jobs in the world today. It used to be this type of work was widely spread throughout programming and technology, but these days many types of jobs are commonly completed using contract workers. I like the flexibility of contract work, but I also think that these workers need to be better at saving and planning for the future because of less employment stability. Usually they are paid more, but they need to use that to reduce the risk of being unemployed.

    One trend that I’ve seen taking place in some of these positions is the use of software and AI to determine if a worker is doing an acceptable level of work on a regular basis. Amazon might be one of the highest-profile companies doing this, especially as they expand into the delivery business. They are using the power of computers to manage an army of workers rather than traditional human managers. This includes terminating them. There’s an article that talks about some of the experiences of their workers.

    I don’t know how their system works, but I do know the frustration of trying to work with a company that doesn’t use humans for many tasks. If you’ve ever tried to contact Google, you know that it’s incredibly difficult to actually communicate with a human. Google seems to think that its automated systems can handle all situations. They might handle many things, but they don’t do a good job in plenty of situations, and there is little recourse to have a human intervene.

    I do think that AI and ML can help our companies better interact with the world in many cases, but these systems are certainly looking for broad patterns. Maybe these patterns handle the middle 80% of cases, or maybe it’s more like 60%, but there are plenty of situations where humans ought to be involved. Maybe more important, when someone uses these systems to make decisions that impact human life, there should be some explanation and understanding of how the model impacts this specific situation. We want to know why the computer comes to its conclusion in medical care, employment, legal issues, or maybe anything other situation.

    There is work being done to try and explain how these models work. The important thing, however, is to ensure that while we may understand the model, we also need to disclose the reasoning to those affected by the systems. Any appeal process should include this explanation, and likely with a human involved at some point to help evaluate the model for accuracy, fairness, or any other measure that is relevant. To me, we ought to require this of companies using AI models in their business practices.

    Steve Jones

    Listen to the podcast at Libsyn, Stitcher, Spotify, or iTunes.

  • Seriously Real Time Data Processing

    There are many of us that work with systems where data is processed in real-time and then used to make decisions. These might be humans viewing reports and then taking action, or some automated system that might react based on a value changing. In many cases, however, the amount of data, timing in which to react with a decision, and the implications for mistakes aren’t that critical. We have some leeway for the processing not being perfect.

    For vehicles traveling at 180mph, the tolerance for mistakes is low, with the chance of a catastrophic crash looming constantly. That is what is happening with a competition at the Indianapolis Motor Speedway, where university students are competing to develop race cars that can move around that track at these speeds. This would be quite a challenge for vehicles moving around the track by themselves, but in this case, it’s a race with multiple cars.

    This might seem silly, but it’s a step towards understanding just how much data needs to be processed and how the results can deal with the chaos in the real world. Each car is independent, so has to react to the other vehicles and make decisions on how to adjust its own operation in real-time, with a sub-second response to prevent accidents. This isn’t different from the decisions human drivers have to make in a race, and there are plenty of mistakes that result in crashes. However, humans can think in new situations and react. They don’t need to have every possible response programmed in.

    These AI-driven race cars will be similar, but how well they perform remains to be seen. This is the type of test environment that will help us move forward in using technology and AI models in less constrained environments, like a public highway. Lots of technology was tried on race tracks before it became available to consumers, and I think this will make its way to retail cars as well at some point.

    There are already some companies trying to build this into cars. Tesla famously has Full Self Driving, although this has been in beta for a long time, and its results are less than stellar in some cases. Waymo has been working on the problem, and I actually had the chance to ride in a self-driving Uber in Las Vegas, though that experience was less than thrilling. The driver had to take control a number of times, so this wasn’t quite self-driving.

    There is a lot of work still to be done here, and I don’t know how quickly this will become safe enough for general driving. I suspect that we’ll see this in very limited areas first, like a zone in a city that only allows these types of cars, or maybe specific highways, like HOV lanes. Somewhere the problem domain is simpler, with less decisions that need to be made.

    There are already lots of safety features in modern cars that help prevent mistakes and accidents, but most of these are simple systems that aren’t making decisions in many ways. Moving to more complex driving operations will require some heavy data gathering, processing, and analysis, something that should be of interest to data professionals. This is a problem domain that will be fascinating to watch in the next few years.

    Steve Jones

    Listen to the podcast at Libsyn, Stitcher, Spotify, or iTunes.

  • Finding Legal Data

    Many of us have likely been asked about data science or machine learning in the last few years by someone in our organization. This has become a hot field, with many companies looking to try and find ways to use the technology to improve their work in some way. While I don’t know how successful these projects have been in organizations, I know that some areas seem to be finding practical applications. Image recognition, translation, and even some application scoring systems have benefited from machine learning algorithms.

    To build a successful model that works well, we need training data. Often lots of training data, and then have some metadata about how our training data might be applicable to a particular question. For example, if we have lots of pictures of dogs, we might need to tag the different breeds in order for a system to differentiate among them. If we want loan applications scored, we should have a corpus of documents that are already scored. This metadata allows the system to learn.

    There is a company, Clearview AI, that markets itself as a facial recognition system. To build their model, they scraped images from YouTube and other Internet sources, without consent from Google or the subjects of the videos. This is interesting, as the data itself is publicly available, but gathering it into a database and using it for other purposes might run afoul of data privacy laws, like the GDPR.

    I don’t quite know how to feel about this use of public data. While I don’t mind people viewing my pictures, I’m not sure that I like the idea of them being copied into a database for some other purpose. That might seem silly, or even strange, but I do think there is power in data and more power in more data. Allowing others to put my images in a database and use them, perhaps to train a model to recognize me, feels like overreach.

    If you need data for your company, or your idea, what can you do? Many people just scrape Google, Facebook, etc., and get data. That might cause you some legal issues in some places, and you ought to be aware of the implications if you choose to do this, or you are asked to do this. I don’t think this is how we want data to be gathered. I know there are some guidelines for responsible AI at Microsoft, but not necessarily rules in place for many companies. Hopefully that will change over time.

    Steve Jones

    Listen to the podcast at Libsyn, Stitcher, Spotify, or iTunes.