AI News, Revolutions


This is an eclectic collection of interesting blog posts, software announcements and data applications I've noted over the past month or so.

The ecosystem for ONNX (the open standard for exchange of neural network models) expands, with official support for Core ML, NVIDIA TensorRT 4, and the Snapdragon Neural Processing Engine.

Oracle acquires machine learning platform, a cloud workspace platform for data science projects and workloads.

Deep neural networks encoded by FPGAs (field-programmable gate arrays) can dramatically reduce the time to classify images (as just one application example).

The TWIML AI podcast series on differential privacy, a technique for collecting data in such a way that it reduces the impact on privacy even in the event of a leak.

This hands-on lab walks through the process of building an image recognizer using transfer learning with MobilenetV1.

How to Develop a Currency Detection Model using Azure Machine Learning, with details on how the real-time banknote recognition capability of the Seeing AI application was implemented in CoreML.

API-driven services bring intelligence to any application

Developed by AWS and Microsoft, Gluon provides a clear, concise API for defining machine learning models using a collection of pre-built, optimized neural network components.

More seasoned data scientists and researchers will value the ability to build prototypes quickly and utilize dynamic neural network graphs for entirely new model architectures, all without sacrificing training speed.

Diving deep into what’s new with Azure Machine Learning

Earlier today, we disclosed a set of major updates to Azure Machine Learning designed for data scientists to build, deploy, manage, and monitor models at any scale.

This post covers the learnings we’ve had with Azure Machine Learning so far, the trends we’re seeing from our customers today, the key design points we’ve considered in building these new features, and dive into the new capabilities.

Before the term was in use, we enabled serverless training of experiments built by graphically composing from a rich set of modules, and then deploying these as a web service with the push of a button.

It has been incredibly rewarding to see how the service has been used by our customers including: Over time we’ve worked with many customers who are looking for the next level of power and control and the capabilities we announced today address those desires.

This demand for AI by developers will only increase, further pushing organizations to provide easy to consume AI built on their data, as the way we write software evolves around these new capabilities.

While we see customers consolidating large amounts of data into data lakes and using tools like Spark for preparing and analyzing their data, the models they produce need to be deployed to a variety of form factors.

Whether looking to address latency or the ability to support scoring even while disconnected, customers want the flexibility to train a model anywhere but control their deployment to place scoring as close to the event as possible.

Over time, this innovation will result in mature toolchains, all the way to the hardware level that are optimized for specific workloads, letting customers tune and tradeoff cost, latency, and control.

Given the learnings we’ve had, we’ve anchored our design on the following four points to shape these new capabilities The services and tools we build must operate at scale, and we see customers encountering challenges with at least five different dimensions of “scale.”

For us, this means embracing container based deployment of models to enable customers fine-grained control, as well as being able to use services such as Azure Container Service to provide a scalable hosting layer in Azure.

Much as source control systems have evolved for software development to flexibly support a variety of teams and processes, our system needs to support the AI development lifecycle as teams continue to grow.

Any service and tool that we build needs to enable data scientists to pick and choose from the ecosystem and use those tools, and we must build it in a way that provides a consistent experience for training, deployment, and management as these evolve.

When we look at the key areas of friction for data science teams, we consistently hear about challenges in: We believe that by eliminating the friction in each of these steps, and between these steps, teams will be able to increase their rate of experimentation.

It’s critical that our customers can have flexibility in their deployment form factor, including: Given these design points, we’ve released the following new capabilities for Azure Machine Learning The Azure Machine Learning Experimentation service allows developers and data scientists to increase their rate of experimentation.

With every project backed by a Git repository, and with a simple command line tool for managing experimentation and training runs, every execution can track the code, configuration, and data that’s used for the run.

More importantly, the outputs of that experiment, from model files, log output, and key metrics are tracked, giving you a powerful repository with the history of how your model evolves over time.

Python libraries from Machine Learning Server (revoscalepy and microsoftml) available with Azure Machine Learning include the Pythonic versions of Microsoft’s Parallel External Memory Algorithms (linear and logistic regression, decision tree, boosted tree and random forest) and the battle tested ML algorithms and transforms (deep neural net, one class SVM, fast tree, forest, linear and logistic regressions).

We know that data science isn’t a linear process, and the Experimentation service lets you look back in time to compare experiments that produced the right results.

Models are exposed via web services written in Python, giving you the ability to add more advanced logic, custom logging, state management, or other code into the web service execution pipeline.

When deploying models at scale on an Azure Container Service cluster, we’ve built a hosting infrastructure optimized for model serving, that handles automatic scaling of containers, as well as efficiently routing requests to available containers.

Retraining scenarios, where a deployed model is monitored and then updated after being trained on new data, are possible, enabling continuous improvement of models based on new data.

The Azure Machine Learning Workbench also hosts Jupyter notebooks that can be configured to target local or remote kernels, enabling iterative development within the notebook on your laptop, or hooked up to a massive Spark cluster running on HDInsight.

We want to reduce the time and effort to acquire data for modeling, and we want to fundamentally change the pace with which data scientists can prepare and understand data, and accelerate the time to get to “doing data science.”

We have combined a variety of techniques, using advanced research from Microsoft Research on program synthesis (PROSE) and data cleaning, to create a data wrangling experience that drastically reduces the time that needs to be spent getting data prepared.

With the inclusion of a simple set of libraries for handling data sources, data scientists can focus on their code, not on changing file paths and dependencies when they move between environments.

By building these experiences together, the data scientist can leverage the same tools in the small and in the large, as they scale out transparently across our cloud compute engines, simply by choosing target environments for execution.

This extension provides a rich set of capabilities for building models with deep learning frameworks including Microsoft Cognitive Toolkit (CNTK), Google TensorFlow, Theano, Keras and Caffe2, while integrating with the Experimentation service for executing jobs locally and in the cloud, and for deployment with the Model Management services.

Data scientists on our team have put together detailed scenario walkthroughs, complete with sample data, for you to get started on some interesting challenges or adapt their techniques to your next problem, including: We’re constantly working on and refreshing the documentation, if you have a comment or suggestion, please let us know.

Enabling AI in your solutions with Azure

Watch this complete tour of AI, machine learning and cognitive services in Microsoft Azure. Artificial intelligence is the process of empowering machines to ...

First look at What’s New in Azure Machine Learning

Take in the huge set of capabilities announced at Ignite for the next generation of the Azure Machine Learning platform. Build and deploy ML applications in the ...

Microsoft Data Platform – SQL Server 2017 and Azure Data Services

All around us, data is driving digital transformation. Companies that invest heavily in cloud, data and AI have nearly double the operating margin of those that do ...

AI for Earth : Analyzing Global Data with Azure : Build 2018

Microsoft has publicly committed $50 million over 5 years for artificial intelligence projects that support clean water, agriculture, climate, and biodiversity. But our ...

Predicting the Winning Team with Machine Learning

Can we predict the outcome of a football game given a dataset of past games? That's the question that we'll answer in this episode by using the scikit-learn ...

New Google AI Can Have Real Life Conversations With Strangers

At its 2018 I/O developer conference, Google showed off some updates coming to Google Home and Assistant. One feature — Google Duplex — can make ...

Amplifying human ingenuity with Microsoft AI

Microsoft is delivering a flexible AI platform for organizations and developers to infuse intelligence into their products and services using tools and services like ...

Microsoft - AI is a Big, Big Deal

Through machine learning, Microsoft has significantly improved its most known products: the Bing search engine, Skype internet phone service, MS Office ...

Intelligent Edge - How 3 lines of code & Windows ML empower devs to run AI locally on Win 10 devices

In part 2 of the 2-part Intelligent Edge series, Rosane and Wilson give an overview of what Windows ML can do for developers; and show how to easily use only ...

IBM Watson: How it Works

Learn how IBM Watson works and has similar thought processes to a human.