Revolutions

AI, Machine Learning and Data Science Roundup: March 2019

A monthly roundup of news about Artificial Intelligence, Machine Learning and Data Science. This is an eclectic collection of interesting blog posts, software announcements and data applications from Microsoft and elsewhere that I've noted over the past month or so.

Open Source AI, ML & Data Science News

TensorFlow Privacy: a Python library for training machine learning models with differential privacy, for use with sensitive data to generate models that don't learn details about specific people.

Tensorflow Federated, an open-source library for Federated Learning, enabling many participating clients to train shared ML models while keeping their data local.

R 3.5.3 has been released.

NNI, an open source AutoML toolkit for neural architecture search and hyper-parameter tuning, from Microsoft Research.

Industry News

Open AI has published a paper describing GPT-2, an unsupervised language model that can generate paragraphs of coherent text that could be mistaken for human writing. Only a scaled-down version has been released, for fear of abuse.

Mozilla releases Common Voices, a public domain dataset of 1,400 hours of recorded voice data in 18 languages.

GCP's Deep Learning Virtual Machine now incorporates RAPIDS, the open-source GPU-acceleration library for Python machine learning.

Google introduces GPipe, an open-source library for training large-scale deep neural networks on devices with otherwise insufficient resources.

Determined AI, platform-agnostic software for automated machine learning and infrastructure management, has been announced.

Microsoft News

Azure Machine Learning Service now incorporates RAPIDS, the open source NVIDIA library that provides accelerated GPU support for Pandas dataframes and scikit-learn machine learning algorithms.

ONNX Runtime adds the NVIDIA TensorRT execution provider, for improved inferencing support on NVIDIA GPU devices.

The Intel Optimized Data Science Data Science Virtual Machine, providing up to 10x performance increase for CPU-based deep learning workloads, is now available on Azure Marketplace.

New Python features in VS Code include validation of breakpoint targets, a new test explorer, and the ability to run selected code without the need to define code cells.

New enhancements for Cognitive Services Computer Vision: bounding boxes for detected objects, and more types of objects detected (including thousands of brand logos).

Azure Cognitive Services containers for Text Analytics and Language Understanding containers are now supported on edge devices with Azure IoT Edge.

MMLSpark 0.16 is released, with improvements to machine learning methods for Apache Spark including Azure Search integration and the Smart Adaptive Recommender.

Microsoft Research releases MT-DNN as a PyTorch package, an architecture for natural language understandings that combines elements of MTL and BERT.

Learning resources

Microsoft launches AI Business School, a collection of learning modules aimed at business leaders on AI strategy, technology and responsible impact.

A high-level article on how and why transfer learning works, and the connection with embeddings.

The Quartz AI Studio, a task-oriented resource for data journalists, provides useful guides for text and image analysis for general practitioners as well.

Causal Inference: The Mixtape, a freely-licensed book by Scott Cunningham that comes with its own playlist.

An architecture for using R at scale on GCP, by Mark Edmondson.

The Python Data Science Handbook (by Jake VanderPlas) in Jupyter Notebook form. Also available in Azure Notebooks and Google Colab.

Datasets for Machine Learning: A list of the biggest datasets from across the web.

Applications

SPADE, a method for synthesizing photo-realistic images from a simple sketch.

Seeing AI now includes a feature to help vision-impaired users explore photographs by touch, and is now supported on iPad devices.

Google deploys a compact RNN transducer to mobile phones that can transcribe speech on-device and streams output letter-by-letter, and a quasi-recurrent neural network for handwriting transcription.

NVIDIA introduces the Kaldi ASR Framework for high-speed speech transcription.

Find previous editions of the monthly AI roundup here.

R 3.5.3 now available

The R Core Team announced yesterday the release of R 3.5.3, and updated binaries for Windows and Linux are now available (with Mac sure to follow soon). This update fixes three minor bugs (to the functions writeLines, setClassUnion, and stopifnot), but you might want to upgrade just to avoid the "package built under R 3.5.4" warnings you might get for new CRAN packages in the future.

R releases typically reference the Peanuts cartoon, but the code-name for this release, "Great Truth", is somewhat of a mystery. There may be a clue in the release date, March 11. Anyone got any ideas?

For more details on this latest update to the R language, check out the announcement below. And as always, thanks to the members of the R Core Team for their contributions to all R users with the R project.

R-announce mailing list: R 3.5.3 is released

 

Where to find the worst weather in the US

Which US city has the worst weather? To answer that question, data analyst Taras Kaduk counted the number of pleasant days in each city and ranked them accordingly. For this analysis, a "pleasant" day is one where the average temperature was in the range 55°F-75°F, the maximum was in the range 60°-90°, the minimum was in the range 40°-70°, and there was no significant rain or snow. (The weather data was provided by the NOAA Global Surface Summary of the Day dataset and downloaded to R with the rnoaa package.)

With those criteria and focusing just on the metro regions with more than 1 million people, the cities with the fewest pleasant days are:

1. San Juan / Carolina / Caguas, Puerto Rico (hot year-round)
2. Rochester, NY (cold in the winter, rain in the summer)
3. Detroit / Warren / Dearborn, MI (cold in the winter, rain in the summer)

You can see the top (bottom?) 25 cities in this list in the elegant chart below (also by Tara Kaduk), which shows each city as a polar bar chart and with one ring for each of the 6 years of data analyzed. 

And if you're wondering which cities have the best weather, here's the corresponding chart for the 25 cities with the most pleasant days. San Diego / Carlsbad (CA) tops that list.

You can find the R code behind the analysis and charts in this Github repository. (The polar charts above required a surprisingly small amount of code: it's a simple transformation of a regular bar chart with the ggplot2 coord_polar transformation — quite appropriate given the annual cycle of weather data.) And for the full description of the analysis including some other nice graphical representations of the data, check out the blog post linked below.

Taras Kaduk: Where are the places with the best (and the worst) weather in the United States

An architecture for real-time scoring with R

Let's say you've developed a predictive model in R, and you want to embed predictions (scores) from that model into another application (like a mobile or Web app, or some automated service). If you expect a heavy load of requests, R running on a single server isn't going to cut it: you'll need some kind of distributed architecture with enough servers to handle the volume of requests in real time.

This reference architecture for real-time scoring with R, published in Microsoft Docs, describes a Kubernetes-based system to distribute the load to R sessions running in containers. This diagram from the article provides a good overview:

You can find detailed instructions for deploying this architecture in Github. This architecture uses Azure-specific components, but you could also use their open source equivalents if you wanted to host them yourself:

• Microsoft Machine Learning Server (and specifically, the mrsdeploy package) is used to expose an R function for prediction as an HTTP endpoint. You can use any method you like to make R into a REST service: for example, standard R and the plumber package.
• Azure Container Registry is where the container images with R and your prediction code are stored. You could also use your own private Docker Registry, or Docker Hub here. 
• Azure Kubernetes Service provides a managed Kubernetes cluster to host the containers and manage them in the underlying server cluster (which can be scaled up and down in size dynamically, according to load). If you have your own Kubernetes cluster, you can substitute it here (but note the security considerations and authentication required).

For more details on this architecture, take a look at the Microsoft Docs article linked below.

Microsoft Docs: Real-time scoring of R machine learning models