Latest Thinking: Scaled Machine Learning

IMG_3015

(T)  This week, I had the opportunity to attend for the third time, the third Scaled Machine Learning conference that was not held at Stanford University but the Computer History Museum in Mountain View. Organized by Professor Reza Zadeh from Stanford University, and his computer vision venture Matroid, the speakers are some of the leading minds in the field, but every talk was quite understandable (at least the concepts).

Following is the introduction for a few selected talks with the video of the presentation:

Domain-specific architectures for deep neural networks –  David Patterson UC Berkeley & Google, Turing Award Winner

David Patterson was a pioneer in the Reduced Instruction Set Computer (RISC) processor architecture in the early 90s with John Hennessy. In this talk, he reviewed the architectures of Google TPU (TensorFlow Processing Unit) first and second generation. Patterson and Hennessy have written a recent article on computer architectures: “A New Golden Age for Computer Architecture“.

 

Accelerating AI  – Bill Dally Stanford University & nVIDIA

Professor Dally as always made a technically interesting presentation outlining nVIDIA’s approach to GPUs for deep learning with many comments regarding the differences between nVIDIA’s GPUs and Google TPUs.

 

 

PyTorch –  Dmytro Dzhulgakov from Facebook AI

Dmytro reviewed the key features of PyTorch 1.0 and its evolution to become a framework for production environment and the merge of Caffe2 and PyTorch.

 

 

AI at scale at Facebook  – Srinivas Narayanan

This presentation from Srinivas is quite interesting as he explained some of the best practices of Facebook to deploy deep learning at scale in production for personalization, language understanding, and computer vision.

 

Inside nVIDIA’s AI Infrastructure for Self-Driving Cars – Clément Farabet

This presentation from Clément introduces nVIDIA deep learning pipeline for self-driving cars, named MagLev, and its development framework for deep learning application on its GPU named RAPIDS.

 

 

GPT-2 from OpenAI – Ilya Sutskever

I was a little bit disappointed by this talk from ILYA about GPT-2 as I was expecting more.

I wrote in a previous blog post an article about GPT-2: OpenAI’s GPT-2 Language Model.

One of the key points of Ilya is the attention mechanism used in the architecture of GPT-2 that leverages Google’s Transformer architecture based on self-attention mechanisms (Attention is All You Need).

 

AI and Security: Lessons and Future Directions – Dawn Song Berkeley & Oasis Labs

A few though at the intersection of AI and security by well-known and a MacArthur recipient Dawn Song:

 

 

Towards a Conscious AI: A Computer Architecture inspired by Neuroscience – Michael Blum

An 80 years old Turing Award researcher Michael Blum from CMU got on a quest to make Turing machines conscious with the simplest model that explains human consciousness in today’s neurosciences…Incredibility interesting and plausible!

 

 

Note: The picture above is from the Misión de San Carlos Borromeo de Carmelo (e.g. the Mission in Carmel).

Copyright © 2005-2019 by Serge-Paul Carrasco. All rights reserved.
Contact Us: asvinsider at gmail dot com.