Last week I attended (and presented) at the Open Data Science Conference (East) in Boston and I was pleasantly surprised to see that mine was not the only presentation that mentioned fastai. While I was only able to attend a small number of the sessions in person, two of them mentioned their use of fastai:
- Combining Millions of Products into One Marketplace using CV and NLP by Ali Vanderveld of ShopRunner was one of the most interesting presentations I attended all weekend. Her team used fastai for both vision and text models, combining the models by concatenating the embeddings from the two networks and then training a simpler model using the embeddings as the inputs.
- Adding Context and Cognition to Modern NLP by Dr. Catherine Havasi from the MIT Media Lab also mentioned using ULMFiT with fastai (this was attended by my colleague, so I don’t have more detail)
It seemed like there was a lot of interest in the topics that utilize ULMFiT. My presentation discussed some of the work we’ve been doing with ULMFiT (posts here and here). Over 150 people attended and I fielded dozens of questions afterwards from people across a range of industries about how they could apply it to their problems.
The conference really hammered home for me that while we (as students of fastai) might take for granted the power of approaches like ULMFiT, to a lot of people it’s a completely new (and interesting!) idea.