QCon AI 2018 - Thoughts and take aways
Last week I attended QCon AI in San Francisco, CA. QCon AI is a branch of QCon focusing on Machine Learning and AI and targeted to software engineers.
In this post I will briefly talk about some of the most interesting talks (for me, of course) and general thoughts and take aways. Most of the conference slides have already been uploaded to the conference schedule page, and I imagine most of the talks will eventually be uploaded into QCon Youtube channel, so go there for details if you’re interested in a particular talk!
Opening keynote
The conference started with a keynote from Uber, Matt Ranney. Matt is a software engineer working in the self-driving car division at Uber. The talk gave an overview of how self-driving cars can help reduce accidents and save time. Specially for Americans, who spend around 293 hours/year in a car! Matt also talked about how a self-driving car works on global terms. He also explained with some level of detail the infrastructure that Uber has for testing and releasing new versions of the cars.
It was a very interesting talk. And no, Matt did not talk about the tragic accident that happened a few weeks ago in Arizona.
The conference
Another very interesting talk was given by Michael Manapat from stripe. The interesting part about that talk was their approach on testing and improving a classification model already deployed in production. If a model is already deployed, how do you keep collecting data for future improvements? If it is already deployed, it is already affecting the outcome. A particular example was detect fraudulent transactions. If you take actions on those transactions, they’re not gonna happen, and thus your future dataset is going to be unbalanced and it is hard to measure the performance of the current model. One thing you can do is to let slip a % of transactions classified as fraudulent and see if they’re actually fraudulent or not. In all the cases, save the prediction and real outcome for testing.
There were many talks about frameworks and visualizations. TensorFlow was a reall star in the conference, and the TensorFlow/TPU presentation by Magnus Hyttsten from Google was simply amazing.
Interpretability was another hot topic, and there were many small talks about tools and techniques to help to explain your models’ results. That’s something ML practitioners don’t think much about until the moment they have to explain the results to someone that has never heard about ML before.
Closing keynote
The conference closed with a very interesting and inspirational talk by Rachel Thomas. Rachel talked about bias in ML and gave some examples. Specially in NLP, where models have been trained with corpus of ordinary texts like emails, articles, etc. Some examples given were black people being misclassified as gorillas, black people being classified as more prone to relapse after being in prison or jobs like “housekeeping” or “nurse” being more tightly associated with women, while words like “business” and “success” being closer to the word man. This is a sign that our training sets are biased, and is something that really matters. Big companies like Google are already taking an action.
Conclusion
Overall a very nice an inspirational conference. Very well organized and talks were generally of high quality. I will definitely attend next year!
