The Friday morning keynotes included a talk about PyTorch by its author Adam Paszke, and “science vs. COVID” from Amy Heineike of Covid19Primer.com.
PyTorch: a modern machine learning research and production platform
Speaker: Adam Paszke
Paszke’s talk focused on how PyTorch is increasingly used in research—many new advancements will appear alongside their PyTorch implementations—but its adoption in industry has been lagging. One of the current goals is to close this gap.
To this end, Paszke talked about some new work in PyTorch:
- Torch Script, which I believe is JIT-compiled PyTorch to make models serializable and generally faster
- Torch Serve, which enables easier mobile deployments
- other things that went super fast
I can tell that this was a good talk, but I don’t believe that I was in its intended audience—I’ve used PyTorch, but not enough to understand enough about what was going on here, or have the upcoming changes be meaningful to me.
Science vs Covid, lessons from Covid19Primer.com
Speaker: Amy Heineike (from Primer, as in Covid19Primer.com)
How was Covid19Primer.com built? It’s built on the quickly growing literature and discourse on COVID-19; they had automated jobs to scrape PubMed, bioRxiv, medRxiv, and arXiv, along with Twitter and newspapers. They used various machine learning methods to make sense of this massive data—topic modeling, classification, unsupervised clustering, and more.
This was hard to summarize—Heineike talked about different things that she had seen in the COVID data, patterns of how researchers talked about data, and how information was propagated.