A TL;DR Recap of Machine Learning Announcements from AWS re:Invent 2019 Keynote #1
Today at Andy Jassy’s Keynote in re:Invent, Amazon has launched a few super interesting ML/AI capabilities (most of which are available starting today and some will be out early this year).
It has become a tradition during these events that while the announcements are great news for AWS customers, there will be quite a few startup companies doing ML infrastructure (including some I know and love) who will have to reinvent themselves following the new products that have been revealed today.
The following is a TL;DR summary of the ML-related announcements from the keynote:
AI-Services Category
1. Amazon Fraud Detector — Upload a file with historically labeled data, choose a model, do some configuration, and get an API that returns a real-time Fraud Score. The model knows how to identify several different types of malicious attempts (account creation, payment attempt, etc.). Interesting service — builds on Amazons vast experience in retail.
2. Amazon Kendra — a search system capable of reading data from a variety of sources (S3, Salesforce, etc.) and produce a natural language query interface on top of it, in order to provide highly concise answers — similar to the Google Search questions feature (“Does the Venetian have free valet?”).
3. Amazon CodeGuru — a really cute service that can connect to GitHub and perform Code Review automatically using ML-based static analysis. Works by recommending best practices, performance enhancements, and bug fixes (e.g. code snippets that work concurrently without locks)
4. Amazon Connect received a crazy set of capabilities called Contact Lens, an analytics system that knows how to analyze customer engagements (calls and texts) and identify features such as “period of time when no one is talking”, “agent talking too fast”, etc., and produce reports/recommendations using this feature space.
On the SageMaker Ecosystem
1. SageMaker Autopilot — an end-to-end AutoML solution (in the style of DataRobot and others): throw a CSV at it and the platform knows how to clean it up a bit, produce basic features, train an order of 50 different models in a distributed environment (including hyperparameter tuning), and recommend a deployment-appropriate machine for inference. Nice Python interface from SageMaker Notebooks — can easily integrate into Jupyter-style work.
2. SageMaker Studio — a rich web IDE for data science to make the development process using notebooks more efficient, centralized and organized. Integrates with all kinds of other nice new services to concentrate work on the model in one place (not just writing models but also running experiments, monitoring, etc.).
3. SageMaker Model Monitor — a production model performance monitoring system (mainly to detect Concept Drift). This service works by integrating with SageMaker’s Inference Endpoint, generating statistics to create a baseline and then monitoring the inference with respect to it. Integrates with CloudWatch to generate alerts (it’s also convenient because you can read logs from the new SageMaker Studio and then monitor the model from there).
4. SageMaker Experiments — a system for conducting and running experiments (models in different configurations, etc.), a product that seems to help a lot in order-making and integrates with the new SageMaker Studio.
5. SageMaker Debugger — All models supported by SageMaker Estimators can now receive a flag asking them to save state during the training phase. Once there is train-output (saved in the form of tensors), you can run a debug job on the debugger-created-tensors and discover interesting things (for example if there is a certain hyperparameter that has received a value that makes the neural net go crazy).
6. Generally speaking, SageMaker Notebooks have received lots of improvements that make them much more similar to Google Colab (a good thing!), including a few basic things that were missing thus far (such as the ability to easily resize the notebook machine and share projects).
