TensorFlow Everywhere North America Summit (2021) Recap

If you didn’t join TensorFlow Everywhere North America, don’t worry I made this summary for you, about what’s new in TF.

6 min readMar 8, 2021

What follows is a recap of the presentation by Laurence Moroney AI Lead at Google in his talk entitled “What’s New in TensorFlow 2.4” part of the “TensorFlow Everywhere North America day 1” held virtually on February 26, 2021. This recap blends direct transcription, insets of Laurence Moroney presentation slides.You can find the recording of the keynote and talks in this link.

What’s New in TensorFlow 2.4!

Within the past few years since launch, TensorFlow and the TensorFlow community has grown into a comprehensive ecosystem of tools for researchers, developers and companies to help them to accomplish their goals, push the state-of-the-art in machine learning and build scalable ML-powered applications.

It has been downloaded over 200 million times and has over 9.7 million contributions from blog articles read and views on YouTube. Over 52 million tutorials and guided views, and more than 800K learners on open online courses.

In order to improve developers experience, the TF team focused on:

Making it easier for software developers as well as AI practitioners.
Giving them the tools they need to be able to understand ML and how they can apply it to their problems.
Helping them to deploy what they learnt.

The overall Ecosystem provided a diverse collection of workflows that will allow practitioners to develop and train models, and to understand how those models were trained. In addition being able to put those models in people’s hands so that they can be useful to them and profitable for ML businesses.

TensorFlow Ecosystem is committed to 2.X:

The over vision is designed for everyone -research, production and deployment- using responsible AI practices.

Cutting edge ML research with TensorFlow 2.X:

In his presentation Laurence Moroney presented some of the newest research papers such as:

Real time sign language detection model using Human Pose Estimation a Unified Text-to-Text Transformer, you can find here.
Massively scaling RL with SEEd RL architecture, you can find here.
RL with Quantum Variational Circuits, you can find here.

Add-ons and extension in the TF ecosystem:

TensorFlow’s growing ecosystem provides new libraries and extensions that help developers accomplish their machine learning goals such as:

TF Probability
TF Graphics
Mesh TensorFlow
TF Model Garden
TF Agents
TF Text
Swift for TensorFlow
Sonnet
Neural Structured Learning
TF Quantum
…. and more on tensorflow.org

Research models migrating to TF 2.X

In this section Laurence Moroney presented some research models migration to TF 2.X such as :

Object Detection API:

TF Recommenders (TFRS) :

To learn more, see the tutorial on how to build a recommender system with examples.

TensroFlow in Production

One of the problems encountered by developers is being able to transition from local debugging to distributed training in Google Cloud, and Laurence Moroney presented a simple code showing how TF team had improved TensorFlow Cloud by simplifying the process of training TensorFlow models on the cloud into a single, simple function call, requiring minimal setup and no changes to the model, also how TF 2.X supports TPUs to make it more easier for developers.

Deployment

TF lite

While working with TensorFlow Lite, as a solution for running models on mobile and embedded systems, Laurence Moroney insists to take different consideration into account:

Lower latency
Network connectivity
Privacy preserving

Tf.js: Train +Deploy ML Models to the browser

TensorFlow.js can be used for building and training machine learning models directly in JavaScript either in a browser or with Node.js support.

Laurence Moroney presented some examples of projects made with TensorFlow.js:

Latest on-device ML with TF Lite

Improvement to TF Lite

What is responsible AI?

The advancement of AI is opening up new possibilities for solving difficult, real-world problems. It also raises new concerns about how to build AI systems that help everyone.

Although every TensorFlow pipeline is likely to face unique challenges and development requirements, TF team noticed a consistent workflow that developers use when creating their own products. Developers are confronted with various Responsible AI questions and considerations at each point of the process. With this workflow in mind, we’re developing our Responsible AI Toolkit to operate in harmony with existing developer processes, allowing Responsible AI efforts to be seamlessly integrated into an existing framework.

To read more about responsible AI you can check TensorFlow Responsible AI and this blog from TensorFlow about Responsible AI with TensorFlow,

Laurence Moroney also mentioned differential privacy, as a way to support privacy which adds noise in the training to hide individual examples in the datasets. TensorFlow Privacy provides a set of optimizers that enable developers to train with differential privacy, from the start.

Model Card Toolkit: Automate, Transparency, Reporting

The Model Card Toolkit (MCT) library are machine learning documents that provide context and transparency into a model’s creation and output. Developers can share their model’s metadata and metrics with researchers, developers, reporters, and others by integrating the Model Card Toolkit into their ML pipeline.

Powered by the Community

Community is, of course, at the core the TensorFlow team interest.You may get involved in a variety of ways as Laurence Moroney mentioned. Consider the following scenarios:

Collaborative ML with TensorBoard.dev
Translation and Localisation: Jupyter notebook community translation support( currently 13 languages), translation guides, tutorials and other documentation into Japanese, Korean, and Simplified Chinese(with more to come)
TensorFlow user Groups: 80 TFUGs around the globe, so, if you want to form one in your area, please reach out at tfug-help@tensorflow.org!
Google Developer experts: 165 ML GDES globally, in 2020, they gave 961 techtalk/workshops and 620 videos/articles, and great collaboration projects like background Stylizer and AI vs COVID
Supporting black AI community: Funding faculty and supporting content at historically black colleges and universities, donating to Black in AI, funding and providing content to AI in Africa chapters and offering certificate stipends and Coursera credits for the above groups.
Certification in TensorFlow Development: over 1K developers from 50+ countries who have passed, the goal was to fill the gap between the employers and the developers so they come up with a set of a syllabus of the commun things that employers are looking for as computer vision, text classification, and Sequences and prediction.