It is always a good idea to attend conferences to know what is new and meet interesting people who work in the amazing Bay area. Here is a summary of events that I attended at the @scale conference this year -

Women’s Leadership Breakfast

At this event the most important thought that stuck with me was think ‘long-term’. Even if you take a break in your career due to some reason, you can always start something new when you come back. This might enable you to provide more value to the company and your team as you look at things from a fresh perspective. A break might enable you to think out of the box instead of being stuck in the rut of delivering code as fast and efficiently as possible. There was a mention of kepler.gl which uses a gpu to analyze huge datasets containing geospatial data in the browser. This will definitely make visualization easier and accessible.

Keynotes

The keynote ‘Golden age for Computer architecture’ mentioned Moore’s law ending but that is not the end of the game. This has lead to a new beginning with domain specific Architectures (DSAs), which are purpose-built processors that try to accelerate a few application-specific tasks. GPUs and TPUs which accelerate neural networks will be the new standard we look up to. Floating point was a big deal in the 1980’s when x86 was introduced. Doing really fast integer based matrix multiplications is the basis of the new architectures.

FB also discussed GLOW (graph-lowering compiler) which will take a neural network graph from Pytorch and then optimize it to run on a ML accelorator. Also ONNX which is an open source format which allows graphs to be converted from one framework like Pytorch to CNTK which is interesting. Pytorch 1.0 was introduced as a combination of the best features of the two — Caffe 2 is optimised to be used in production (is scalable and fast) and Pytorch which is used in research.

Nvidia discussed Project Maglev for autonomous driving where K8s is used for programatic workflows on GPU clusters. Another interesting thing mentioned was the TensorRT inference server. Developers can focus on creating models instead of optimising performance for deployed models. RTCores were introduced at SIGGRAPH this year as a part of Turing, which make real-time ray tracing possible. The company introduced shiny new gpu’s Quadro RTX 8000/6000/5000 and Quadro RTX Server for datacenters. Also at Gamescom this year they announced Geforce RTX 2080Ti/2080/2070.

There was an interesting demo by FB for Mask R-CNN that tracks human movements in real time. There was a demo by Microsoft — Brainwave which is a FPGA powered DNN serving platform which works today with TF and Pytorch support coming soon. There were interesting demos for Nvidia HGX-2 and sentiment detection.

Friends Don’t Let Friends Deploy Black-Box Models: The Importance of Intelligibility in Machine Learning

The more complex ML models are the harder they are to understand. Boosted trees, RF, Neural nets are harder to make sense of than Linear regression, Logistci Reg, NB and Decision lists models. This talk was very well delivered and mentions how Generalized additive modes (GAMs) attempt to solve this problem.

Applied Machine Learning at Facebook: An Infrastructure Perspective

Quick overview of how ML is used at FB. SVM for Facer, GBDT for Sigma, Multi Layer perceptron MLP for news feed, ads, search and sigma. CNN for Facer and LUMOS, RNN for language translation, speech recording and content understanding. Also covered how often models are trained, how much compute is required for each type of workload and how volta GPUs are used. Also ONNX which is a shared models and operator representation.

Machine Learning Testing at Scale

This talk mentioned how testing is important after model deployment and the opportunities for testing at every step like preprocessing etc.

Computer Vision at Scale as Cloud Services

Rapid rise of DL - Alexnet, Googlenet and Resnet. Microsoft covers how they use computer vision. How privacy was a concern for face and gender detection. Challenges like how the only found “happy”faces on the net and could’nt find many sad, shock or contempt type of face images. They had to retrain their model after generating enough of these type of images.

Scaled Machine Learning Platform at Uber

Described ‘Michalangelo’ an internal machine learning platform that supports Big data at Uber.

That wraps it up!