Fastai v3… what has (not) changed
Jeremy Howard and Rachel Thomas launched yesterday (10/22/2018) the version 3 of their Fastai course on Deep Learning. Like thousands of other people around the world (see the map below), I’m following this course through live video on YouTube.
As I followed the version 2 of the course last year, I’ve appreciated in the course of yesterday the differences (use of the fastai v1 and new notebooks) but also the constancy in the objectives and pedagogy.
What has not changed
- The fastai library is an open source Deep Learning framework developed on PyTorch.
- fastai integrates the latest algorithms on DL and public databases.
- The goals of fast.ai (see the screenshot below) are — based on the fastai library — to participate in DL research, to animate a community and to train as many people as possible in Deep Learning (without going through the PhD box).
- The objective of competence for the participants of this course (see the screenshot below) is to know how to create world class models of DL: a classifier of photos, a sentiment analyzer of texts, an algorithm of predictions (of sales) and an algorithm of recommendations (of films).
- The course pedagogy follows the same Top-Bottom methodology: knowing how to train a model and understand the principles before learning the mathematical and programmatic theory that underlies them.
- Finally, fast.ai wants to fight the “fake ideas” on Deep Learning (see the screenshot below).
What has changed
If you are not registered in the course which has just begun, it will be necessary to wait until early 2019 to know it ;-) but the main change concerns the fastai framework which has just been published in v1 (October 2018) based on PyTorch v1.