A new nest in the same old tree. Twitter have finally migrated to Tensorflow as their preferred choice of machine learning framework. While not many are surprised by this move given the popularity of Tensorflow, many have surely asked the question – ‘What took them so long?’
Why Twitter migrated to Tensorflow only now
Ever since its inception, Twitter have been using their trademark internal system called as DeepBird. This system was able to utilize the power of machine learning and predictive analytics to understand user data, drive engagement and promote healthier conversations.
DeepBird primarily used Lua Torch to power its operations. As the support for the language grew sparse due to Torch’s move to PyTorch, Twitter decided it was high time to migrate DeepBird to support Python as well – and started exploring their options. Given the rising popularity of Tensorflow, it was probably the easiest choice Twitter had to make for some time.
Per the recently conducted Stack Overflow Developer Survey 2018, Tensorflow is the most loved framework by the developers, with almost 74% of the respondents showing their loyalty towards it. With Tensorflow 2.0 around the corner, the framework promises to build on its existing capabilities by adding richer machine learning features with cross-platform support – something Twitter will be eager to get the most out of.
How does Tensorflow help Twitter?
After incorporating Tensorflow into DeepBird, Twitter were quick to share some of the initial results. Some of the features that stand out are:
- Higher engineer productivity – With the help of Tensorboard and some internal data viz tools such as Model Repo, it has become a lot easier for Twitter engineers to observe the performance of the models and tweak them to obtain better results.
- Easier access to Machine Learning – Tensorflow simplified machine learning models which can be integrated with other technology stacks due to the general-purpose nature of Python.
- Better performance – The overall performance of DeepBird v2 was found to be better than its predecessor which was powered by Lua Torch.
- Production-ready models – Twitter plan to develop models that can be integrated to the workflow with minimal issues and bugs, as compared to other frameworks such as Lua Torch.
With Tensorflow in place, Twitter users can expect their timelines to be full of relatable, insightful and high quality interactions which they can easily be a part of. Tweets will be shown to readers based on their relevance, and Tensorflow will be able to predict how a particular user will react to them.
A large number of heavyweights have already adopted Tensorflow as their machine learning framework of choice – eBay, Google, Uber, Dropbox, and Nvidia being some of the major ones. As the list keeps on growing, one can only wonder which major organization will be next on the list.