CliffsNotes + commentary on Jeff Dean’s TensorFlow talk

Li Jiang
Li Jiang
Apr 10, 2017 · 4 min read

This talk by Jeff Dean, Rajat Monga, and Megan Kacholia is about the launch of TensorFlow, an open-source machine learning (ML) platform by Google.

Li: this talk is primarily for a technical audience and is a brief update on the progress of TensorFlow. You should try other videos if you want to learn more foundational knowledge on ML.

I encourage you to watch the full talk if you plan on building with TensorFlow (30 mins, link at end of this post), but if you are just curious or want a 3-5 minute preview, here we go:

part 1 — Jeff (0:05):

Machine learning will be very important in many industries.

We wanted to build something that was open sourced and that was

  1. Fast
  2. Flexible
  3. Production Ready

We built DistBelief in 2012 that could work with very scalable and powerful models in areas including speech recognition, search, image recognition, but it wasn’t exactly what we wanted. It wasn’t as flexible; it was designed for CPU, but GPU support wasn’t added until later.

TensorFlow will be more flexible for research and production; it will run on CPUs, GPUs, Tensor Processing Unit (TPUs), Android, iOS, Raspberry Pi. It also runs well on the cloud and supports many languages. We’ve included tools like TensorBoard for visualization.

Progress to date:

We started in Q1 2012 with nearly no project directories with ML within Google. Now in Q4 2016 we have 4,000 unique project directories with ML.

TensorFlow is the #1 repository in “machine learning” on Github.

Nearly 500 non-Google contributors, 14,000 commits in 14 months. There are now 5,500 Github repositories with TensorFlow, 5,000 Stack Overflow questions, and adoption in curriculums at Toronto, Berkeley, Stanford.

All of this has happened in just over a year.

part 2 — Rajat (7:28):

We are announcing today TensorFlow 1.0. It is fast, flexible, production-ready, and ensures backwards compatibility.

New high-level APIs should make deep learning accessible to everyone.

Li: when they say “everyone”, I think they meant every computer scientist and engineer, ha!

On top of the Python frontend, we are adding Layers, Keras integration, Estimator API. There is broad machine learning support — K-means, Support Vector Machines, Random Forest, etc. It’s integrated with new platforms and accelerators — TensorFlow is included in IBM’s PowerAI distribution, Movidius Myriad 2 accelerator, Qualcomm Hexagon DSP.

Travis Lanier, senior product managing director at Qualcomm: 8x performance increase by moving from CPU to Hexagon DSP — we’ve done the optimization for you and we are open sourcing this software and driver.

Today we are announcing XLA: An Experimental TensorFlow Compiler. XLA = accelerate linear algebra.

part 3 — Megan (15:31):

Performance matters for research — TensorFlow allows you to iterate quickly, train models faster, run more experiments in parallel

On the production side —you can run TensorFlow on large scale server farms and embedded on devices, CPUs, GPUs, TPUs and low-latency serving.

Cutting edge research:

  1. Neural machine translation — TensorFlow reduces errors by 55%-85%.
  2. Neural architecture search —figure out what is the right neural network to use to solve a problem.
  3. Show and tell — provide human sounding captions for images.
  4. ML in production — a lot of production in apps such as Word Lens.

The question for TensorFlow is how we give this back to the community.

part 4 — Jeff again (21:45):

TensorFlow is really for everyone.

Some of the things people have done with TensorFlow are things that we’ve never thought of.

TensorFlow is used to diagnose diabetic retinopathy, the leading cause of blindness in the world.

The Journal of American Medical Association: a computer vision model can be as good or slightly better than the median opthamologist. The world doesn’t have enough access to opthamologist so this can really improve access.

Brett Coprell, Stanford — classifying images of melanoma and skin cancer at dermatologist level performance. Check out the cover article of Nature.

There are a lot of creative uses of TensorFlow, even building a painting robot or making music. Thank you.

Watch the full video:

Li Jiang

Written by

Li Jiang

Business Development @

Welcome to a place where words matter. On Medium, smart voices and original ideas take center stage - with no ads in sight. Watch
Follow all the topics you care about, and we’ll deliver the best stories for you to your homepage and inbox. Explore
Get unlimited access to the best stories on Medium — and support writers while you’re at it. Just $5/month. Upgrade