Week 11, 2019 - TensorFlow Developer Summit; Thoughts on Machine Learning

The TensorFlow Developer Summit was held and has some interesting new announcements. I also mention some of my own thoughts on the idea of machine learning.

TensorFlow Developer Summit

The latest developer summit for TensorFlow was held and a lot of new things came out of that. As a quick recap, TensorFlow is an open source, Google developed, machine learning platform. In fact, it is one of the most popular and powerful machine learning platforms and basically consists of an end-to-end solution.

From my reading of what came out of this event is that they’re working hard on making TensorFlow more accessible and usable for beginners. Personally I love seeing that, as I’ve found in general that when tools are made easier for beginners, it not only ensures more people will use it but it also enables more advanced users to pick up new tricks.

In this case, the main effort seems to be with the introduction of TensorFlow 2.01, but also by extending language support. For JavaScript developers they released TensorFlow.js 1.02, which helps them to build machine learning models in a language they’re familiar with, and they even have a way to use it with Swift. This last one is especially interesting in combination with the release of TensorFlow Lite 1.0, the mobile optimised version of TensorFlow as it might potentially allow you to use all that power in your iOS devices more effectively than with the JavaScript version3. There are a lot more announcements, but it’s probably best to read the article at the top to find out what is useful for you.

Thoughts on Machine Learning

Because while I love all of the progress in the machine learning space, I do wonder how much of this is something most developers will use4. Yes, there are situations where the ability to train your application is important5, but for many applications I wouldn’t be surprised if instead they will start relying on external services to do that for them.

If you want to use some kind of image recognition, would you build your own or use one that is provided by a major cloud provider6? Yes, there might be reasons why it’s worth investing a lot of money and effort into building your own, but in most cases that’s not really the case. Actually, I wouldn’t be surprised if most of the arguments will be repeats of these for using cloud computing in the first place. Or serverless compute. Or… I’m sure you can come up with some more examples.

Again, don’t get me wrong, I think ML is important and that making it easier to build your own solutions is a great step forward. However, in the not very long run, I do believe that a lot of it will become part of ready-made solutions that can be accessed via an API and where you optionally provided some training data7.

On a Personal Front

You probably noticed that the past weeks, over a month actually, lacked anything new appearing on this site. There are reasons for this. One of these is that I didn’t really manage to write anything that I thought was good enough. While I sometimes repeat some of what the press releases say, I do try to give my own opinion on them. After all, what would be the point of writing it if I don’t add anything? You could just read those press releases yourself.

Underlying this however may be some changes for me personally that distracted me. I’ve spent nearly three years at Bulletproof, but late January I believed that it was time to look for a new challenge and decided to join DigIO as Lead Platform Engineer. As you can imagine, this means that most of February I was busy tidying up matters at Bulletproof and ensuring everything was handed over properly.

After that I took a break by traveling a bit through Western Australia. Where I still am right now, and which is the reason for the different posting date. As I’ll be in rural areas for the next days I don’t know if I’ll have an internet connection8, so it’s going up while I have the chance.


  1. Obviously now in alpha, I did say it was a Google product after all. [return]
  2. As a side note here, the graph in there that shows the speed improvements is impressive. Leaving aside the huge gap between Android and iOS, TensorFlow.js is now faster on an iPhone X than it was last year on a MacBook Pro with a dedicated GPU. [return]
  3. Let’s be clear, especially the latest iPad Pros are ridiculously overpowered for the software running on it. I’d love to see more software that can use all of this power, and if this opens that up that’s great news. [return]
  4. Be warned, the below is not completely formed in my head and I might solidify my ideas around this more later. [return]
  5. Even in gaming, I think it would be a lot of fun to see a strategy game where the computer has actually learned from what you did to it in the past. Until it becomes impossible to beat of course. [return]
  6. As a not 100% accurate example, compare Google Photos which does all its machine learning in the cloud with Apple’s Photos which does it on your device. [return]
  7. Of course, you might just be that special case where it makes sense to make a big investment before you know if it will work. Instead of trying the out of the box solution and then decide if you need to make improvements. [return]
  8. I often forget how big and sparsely populated this country is and what that means for things I usually take for granted. [return]
comments powered by Disqus