Commenting on Software Technology Trends & Market Dynamics

Jnan Dash

Subscribe to Jnan Dash: eMailAlertsEmail Alerts
Get Jnan Dash: homepageHomepage mobileMobile rssRSS facebookFacebook twitterTwitter linkedinLinkedIn

Related Topics: cloudbigdataajaybudhraja

Blog Feed Post

The ongoing revolution in machine learning

Machine learning is moving rapidly into the mainstream. Google folks have done a great job in building their framework TensorFlow to enable machine learning development. During a recent event in July called CloudNext’18 Google outlined several enhancements to TensorFlow.

First of all, what is machine learning (ML)? The classical programming we all grew up with taught us to feed data and rules into this black box called the program and out came the answers. We used the if-then-else logic applying the rules to get to the right answer. Machine learning flips this model: we feed data and answers to the black box and outcomes the rules on the other end. In other words, by feeding varieties of answers, we build rules which get upgraded or refined continuously. For example, how do we build an activity tracker to distinguish between walking and running? How do we do face detection, or landmark identification, or explicit content detection, or label identification, etc.? In the era of huge amounts of data we can build inference engines via machine learning.

TensorFlow is a machine learning platform for everyone to solve real problems. Google has made it open source and anyone can do a free download and start experimenting. Here are the actual examples of its use:

  • Haunting for new planets – TensorFlow was used last year to discover new details on the planetary system called Kepler 90. For example, it found out the Kepler 90i planet revolves in 14 days (365 days for us in the earth).
  • Diabetic retinopathy – by measuring the blood pressure effects on the retina muscles, doctors can prevent blindness with early detection.
  • Deforestation – save forests by alerting authorities to signs of illegal deforestation activities. It uses sounds of trees falling or chainsaw working to send alerts.
  • AlphaGo and Google Cloud Vision are built on top of TensorFlow.

During last 12 months, Google has done many improvements to TensorFlow for wider use by regular programmers. Here are some highlights:

  • Support of many languages like Python, C++, C#, Go, R, Julia, Java, Haskell, a JS (JavaScript). For javascript developers, the TensorFlow can be run in the browser for training and execution (no round trips needed to Google cloud or AWS).
  • Introduction of TensorFlow Lite to run inference on mobile and IoT devices.
  • Using TensorFlow.js, you can train and execute Models in the browser. You can use Keras to build neural networks very easily (line by line debugging, easy prototyping).
  • Google has introduced a specialized chip called TPU (T for Tensor) which is highly optimized for running big data and models. It runs much faster than GPUs and costs quite a bit less. One big bottleneck has been the cost of modeling which runs into thousands of dollars. TPU processing benchmark shows it can be done in tens of dollars.

Machine learning is coming out of the academic nerd-land that demanded very high skills. With tools like TensorFlow Hub (repository of pretrained model components) developers can gain rapid productivity with their favorite language like javascript or C++.

Google has done a great job here.

Read the original blog entry...

More Stories By Jnan Dash

Jnan Dash is Senior Advisor at EZShield Inc., Advisor at ScaleDB and Board Member at Compassites Software Solutions. He has lived in Silicon Valley since 1979. Formerly he was the Chief Strategy Officer (Consulting) at Curl Inc., before which he spent ten years at Oracle Corporation and was the Group Vice President, Systems Architecture and Technology till 2002. He was responsible for setting Oracle's core database and application server product directions and interacted with customers worldwide in translating future needs to product plans. Before that he spent 16 years at IBM. He blogs at