No one wants to host conferences in the summer because schools out and it’s time to hit the road with the family. And you can’t expect anyone to attend a conference between the end of November and the middle of January since there are so many holidays and vacations. So that relegates spring and fall to be the official conference seasons, and it seems to get more packed every year. Just look at how the number of conferences with artificial intelligence (AI) in the title is exploding. So, we have to be selective, recognize you can’t do the whole circuit and look for “trip reports” from the ones you missed. This series is our small contribution to the AI community that is trying to keep up on conference highlights. No endorsements or recommendations just some hopefully useful observations.
Dell EMC had both a booth and a great turn out of experts onsite for the OReilly Artificial Intelligence Conference held Sept 9-12 in San Jose CA. Ramesh Radhakrishnan, a Distinguished Engineer with Dell Technologies, John Zedlewski from NVIDIA® had a session titled From inception to insight: Accelerating AI productivity with GPUs. They talked about how a simple set of technologies from Dell and NVIDIA can provide the flexibility to easily work in multiple data science development environments. The key is to leverage NVIDIA GPU Cloud containers on a wide range of validated platforms from laptops to workstations to full-blown enterprise-class compute clusters. Check it out using the link above.
I spent Tuesday attending the Intel® AI Builders Showcase Event: AI in the Enterprise. I spoke to attendees about the Dell EMC work with the Nauta open source software recently released to open source by Intel. This was a very technical group of hands-on developers. It reminded me of the experience I had at the Intel AI Devcon in that I attended in SF last year. Lots of great conversations after every talk.
There were two sessions that I attended this year at the AI Conference in SJ that I want to share with you:
Ankur Taly, Head of Data Science at Fiddler, spoke on Explaining Machine Learning Models. Ankur is well-known for his contribution to developing and applying Integrated gradients - a new interpretability algorithm for deep neural networks. This session was a mental work out but fortunately, I was just reading about Shapley values for interpreting predictions from machine learning models. There are too many articles of varying levels of technical depth, so I suggest you just do a web search if interested. Ankur described a method of attribution called integrated gradients that I had not heard of before. It is applicable to a variety of deep neural networks including object recognition, text categorization, machine translation and more. The method can be used to debug model predictions, increase model transparency, and assess model robustness. He also talked about Shapley values that have lately been applied to explaining predictions made by nondifferentiable models such as decision trees, random forests, gradient-boosted trees.
Alex Ratner, project lead of Snorkel, gave a talk titled Building and managing training datasets for ML with Snorkel. Snorkel is a set of tools that includes labeling functions for labeling unlabeled data, transformation functions for expressing data augmentation strategies, and slicing functions for partitioning and structuring training datasets. There are many use cases where the cost of manual labeling, especially when it requires experts, is so high the machine learning project cannot get started. The Snorkel toolset and methodology are very innovative and have already solved some otherwise intractable labeling challenges. Check it out when you can.
Thanks for reading
Phil Hummel - Twitter @GotDisk