Don’t Build Big Data With Bad Data

I was at Pure Accelerate 2017 this week and I saw some very interesting things around big data and the impact that high speed flash storage is going to have. Storage vendors serving that market are starting to include analytics capabilities on the box in an effort to provide extra value. But what happens when these advances cause issues in the training of algorithms?

Garbage In, Garbage Out

One story that came out of a conversation was about training a system to recognize people. In the process of training the system, the users imported a large number of faces in order to help the system start the process of differentiating individuals. The data set they started with? A collection of male headshots from the Screen Actors Guild. By the time the users caught the mistake, the algorithm had already proven that it had issues telling the difference between test subjects of particular ethnicities. After scrapping the data set and using some different diverse data sources, the system started performing much better.

This started me thinking about the quality of the data that we are importing into machine learning and artificial intelligence systems. The old computer adage of “garbage in, garbage out” is never more apt today than it has been in history. Before, bad inputs caused data to be suspect when extracted. Now, inputting bad data into a system designed to make decisions can have even more far-reaching consequences.

Look at all the systems that we’re programming today to be more AI-like. We’ve got self-driving cars that need massive data inputs to help navigate roads at speed. We have network monitoring systems that take analytics data and use it to predict things like component failures. We even have these systems running the background of popular apps that provide us news and other crucial information.

What if the inputs into the system cause it to become corrupted or somehow compromised? You’ve probably heard the story about how importing UrbanDictionary into Watson caused it to start cursing constantly. These kinds of stories highlight how important the quality of data being used for the basis of AI/ML systems can be.

Think of a future when self-driving cars are being programmed with failsafes to avoid living things in the roadway. Suppose that the car has been programmed to avoid humans and other large animals like horses and cows. But, during the import of the small animal data set, the table for dogs isn’t imported for some reason. Now, what would happen if the car encountered a dog in the road? Would it make the right decision to avoid the animal? Would the outline of the dog trigger a subroutine that helped it make the right decision? Or would the car not be able to tell what a dog was and do something horrible?

Do You See What I See?

After some chatting with my friend Ryan Adzima, he taught me a bit about how facial recognition systems work. I had always assumed that these systems could differentiate on things like colors. So it could tell a blond woman from a brunette, for instance. But Ryan told me that it’s actually very difficult for a system to tell fine colors apart.

Instead, systems try to create contrast in the colors of the picture so that certain features stand out. Those features have a grid overlaid on them and then those grids are compared and contrasted. That’s the fastest way for a system to discern between individuals. It makes sense considering how CPU-bound things are today and the lack of high definition cameras to capture information for the system.

But, we also must realize that we have to improve data collection for our AI/ML systems in order to ensure that the systems are receiving good data to make decisions. We need to build validation models into our systems and checks to make sure the data looks and sounds sane at the point of input. These are the kinds of things that take time and careful consideration when planning to ensure they don’t become a hinderance to the system. If the very safeguards we put in place to keep data correct end up causing problems, we’re going to create a system that falls apart before it can do what it was designed to do.


Tom’s Take

I thought the story about the AI training was a bit humorous, but it does belie a huge issue with computer systems going forward. We need to be absolutely sure of the veracity of our data as we begin using it to train systems to think for themselves. Sure, teaching a Jeopardy-winning system to curse is one thing. But if we teach a system to be racist or murderous because of what information we give it to make decisions, we will have programmed a new life form to exhibit the worst of us instead of the best.

Advertisements

Flash Needs a Highway

CarLights

Last week at Storage Field Day 10, I got a chance to see Pure Storage and their new FlashBlade product. Storage is an interesting creature, especially now that we’ve got flash memory technology changing the way we think about high performance. Flash transformed the industry from slow spinning gyroscopes of rust into a flat out drag race to see who could provide enough input/output operations per second (IOPS) to get to the moon and back.

Take a look at this video about the hardware architecture behind FlashBlade:

It’s pretty impressive. Very fast flash storage on blades that can outrun just about anything on the market. But this post isn’t really about storage. It’s about transport.

Life Is A Network Highway

Look at the backplane of the FlashBlade chassis. It’s not something custom or even typical for a unit like that. The key is when the presenter says that the architecture of the unit is more like a blade server chassis than a traditional SAN. In essence, Pure has taken the concept of a midplane and used it very effectively here. But their choice of midplane is interesting in this case.

Pure is using the Broadcom Trident II switch as their networking midplane for FlashBlade. That’s pretty telling from a hardware perspective. Trident II runs a large majority of switches in the market today that are merchant silicon based. They are essentially becoming the Intel of the switch market. They are supplying arms to everyone that wants to build something quickly at low cost without doing any kind of custom research and development of their own silicon manufacturing.

Using a Trident II in the backplane of the FlashBlade means that Pure evaluated all the alternatives and found that putting something merchant-based in the midplane is cost effective and provides the performance profile necessary to meet the needs of flash storage. Saturating backplanes with IOPS can be accomplished. But as we learned from Coho Data, it takes a lot of CPU horsepower to create a flash storage system that can saturate 10Gig Ethernet links.

I Am Speed

Using Trident II as a midplane or backplane for devices like this has huge implications. First and foremost, networking technology has a proven track record. If Trident II wasn’t a stable and reliable platform, no one would have used it in their products. And given that almost everyone in the networking space has a Trident platform for sale, it speaks volumes about reliability.

Second, Trident II is available. Broadcom is throwing these units off the assembly line as fast as they can. That means that there’s no worry about silicon shortages or plant shutdowns or any one of a number of things that can affect custom manufacturing. Even if a company wants to look at a custom fabrication, it could take months or even years to bring things online. By going with a reference design like Trident II, you can have your software engineers doing the hard work of building a system to support your hardware. That speeds time to market.

Third, Trident is a modular platform. That part can’t be understated even though I think it wasn’t called out very much in the presentation from Pure. By having a midplane that is built as a removable module, it’s very easy to replace it should problems arise. That’s the point of field replaceable units (FRUs). But in today’s market, it’s just as easy to create a system that can run multiple different platforms as well. The blade chassis idea extends equally to both blades and mid or backplanes.

Imagine being able to order a Tomahawk-based controller unit for FlashBlade that only requires you to swap the units at the back of the system. Now, that investment in 10Gig blade connectivity with 40Gig uplinks just became 25Gig blade connectivity with 100Gig uplinks to the network. All for the cost of two network controller blades. There may be some software that needs to be written to make the transition smooth for the consumers in the system, but the hardware is more than capable of supporting a change like that.


Tom’s Take

I was thrilled to see Pure Storage building a storage platform that tightly integrates with networking the way that FlashBlade does. This is how the networking stack is going to be completely integrated with storage and compute. We should still look at things through the lens of APIs and programmability, but making networking and consistent transport layer for all things in the datacenter is a good start.

The funny thing about making something a consistent transport layer is that by design it has to be extensible. That means more and more folks are going to be driving those pieces into the network. Software can be created on top of this common transport to differentiate, much like we’re seeing with network operating systems right now. Even Pure was able to create a different kind of transport protocol to do the heavy lifting at low latency.

It’s funny that it took a presentation from a storage company to make me see the value of the network as something agnostic. Perhaps I just needed some perspective from the other side of the fence.