Revolutions

AI, Machine Learning and Data Science Roundup: July/August 2019

A mostly 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

StanfordNLP: a pure-Python package for grammatical analysis of sentences in over 50 languages.

New projects added to the PyTorch ecosystem: Skorch (scikit-learn compatibility), botorch (Bayesian optimization), and many others.

PyTorch-Transformers, a library of pretrained NLP models (BERT, GPT-2 and more) from HuggingFace.

R 3.6.1 is released.

Announcing mlr3, a new machine-learning framework for R.

Dash for R: a framework for building interactive web applications on both Python and R models.

R now supports the Apache Arrow data interchange format via the arrow package on CRAN.

A Nature article about Julia, the scientific computing language.

Industry News

Ethical Aspects of Autonomous and Intelligent Systems,  the position statement on AI ethics from the IEEE.

ParlAI, a Python framework and repository of data sets and reference models which Facebook is using to address weaknesses in chatbots today.

Facebook releases ELI5, a set of scripts to download paired questions and answers from the ELI5 subreddit, along with models for question-answering research.

Facebook open sources CraftAssist, a platform for creating AI assistants that can manipulate the Minecraft world via chat conversations.

Lyft releases the Level 5 Dataset, the largest publicly-released dataset for autonomous driving models.

Uber releases the Plato Research Dialogue System, a framework for creating, training, and evaluating conversational AI agents.

Uber releases Ludwig v0.2. This update to the code-free deep learning toolbox adds support for Comet.ml and BERT models, and a deployment server.

EfficientNet-EdgeTPU, a family of image classification models optimized to run on Google's low-power Edge TPU  chips.

The What-If Tool is now available within GCP AI Platform, to visualize the impact of variables on Tensorflow model outputs.

IBM introduces AI Explainability 360, a suite of open-source tools for machine learning interpretability.

Microsoft News

Microsoft invests $1B in Open AI to spur research in general artificial intelligence.

AI for Cultural Heritage, a new pillar in Microsoft's $125M AI for Good portfolio.

VS Code adds support for debugging Python cells in Jupyter Notebooks.

Azure Form Recognizer (preview) adds the ability to extract structured data from scanned receipts.

Cognitive Services Text Analytics can now provide sentiment scores for sentences and entire documents, and with improved accuracy.

Video Indexer now allows the language model to be customized, so that specialized vocabulary (jargon) can be automatically detected and transcribed.

Microsoft Machine Learning Server 9.4 is now available, with updates to the open-source engines for operationalizing R and Python.

AzureR, a family of packages for creating, managing and monitoring Azure services from the R language.

Azure Machine Learning Services adds a command-line interface complementing the existing Python SDK, to support MLOps workflows via the CLI.

Microsoft open-sources scripts and notebooks to pre-train and finetune the BERT natural language model with domain-specific texts.

Learning resources

A beginner's introduction to recurrent neural networks from Victor Zhou, with a from-scratch implementation of a sentiment analysis RNN in Python.

Two new, free courses from fast.ai: Deep Learning from the Foundations and A Code-First Introduction to Natural Language Processing.

Statistics with Julia, a draft (but comprehensive) book on statistical analysis in the Julia language.

Keynote presentations from the useR!2019 conference. Contributed presentations from the annual R conference coming soon.

Model interpretability with automated machine learning in Azure Machine Learning service.

Tutorial: Deploying Azure ML Service models to Azure Functions for inference.

A comparison of pre-trained text-recognition services from Amazon, Google and Microsoft.

What's the difference between deep learning, machine learning, and AI?

A tutorial on pre-training BERT models with Google Cloud TPUs.

Debugging ML models: dealing with information leakage and data bias.

Applications

Stitch Fix has built a centralized experimentation platform, to standardize and automate the process of A/B testing.

IMT Atlantique, a Microsoft AI for Earth grantee, uses satellite remote sensing data to monitor changes in ocean levels.

How AI is helping track endangered species.

The Climate Modeling Alliance uses Julia and CUDA to simulate climate models.

Parrotron: a Google Research project to facilitate understanding of speech-impaired people by devices and other people.

Find previous editions of the monthly AI roundup here.

Data on demand: Azure SQL Database in serverless mode

Azure SQL Database has a new “serverless” mode in preview that eliminates compute costs when not in use. In this post, I’ll show how you can set up a serverless database instance, and access data stored in it from R.

I’m working on a demo that I’ll be giving at several upcoming conferences, and for which I’ll be needing data in a database. Normally, I’d use a database installed on my local machine or in a virtual machine in the cloud, but this time I decided to go a different route: serverless.

I need the database to be around for a few months, but I’ll only be accessing it occasionally while I develop the demo and, later, present it live a few times. I need the database to be fairly large and fast (both of which rule out installing it on my laptop), and I’d prefer not to have to pay so much for the cloud resources while I’m not using it. (Yes, as a Microsoft employee I do have access to Azure services, but our department is cross-charged for our consumption and our spend is scrutinized like any other expense.) I considered using specialized cloud-data services, but for this talk I needed something that looked and felt like a traditional database.

I could always run a database in a virtual machine, but this time I decided to try Azure SQL Database. Azure SQL Database runs just like an ordinary SQL Server instance on a named server, but without the need to create or manage any associated VM. Even better, I discovered that it now has a “serverless” pricing tier, which gets me several benefits: