jph00

48 ms·

661.
by jph00 8y ago
You'd almost certainly be better off with the deep learning course ( http://course.fast.ai ) then, not the machine learning course discussed in this post. I spent quite a bit of time in medical imaging, and lots of radiologis
662.
by jph00 8y ago
> I actually wrote most of the machine learning content at Dataquest (where I work) and I started with k-nearest neighbors because it's way more approachable That's a great idea. It's actually what I did in an earlier vers
663.
by jph00 8y ago
Due to our teaching and research cycle, we do generally release a bunch of related things about once a quarter, around when new courses come out. So generally you won't see any fast.ai related news for a few months, then a few things m
664.
by jph00 8y ago
Yes I do think his ML course is still wonderful, as I said in the post :) However I don't think that Octave is better than Python for teaching or understanding machine learning concepts.
665.
by jph00 8y ago
Andrew's new deep learning courses ( http://www.deeplearning.ai ) all use Python. When I last spoke to him he was no longer enthusiastic about using Octave/Matlab for teaching.
666.
by jph00 8y ago
v3 of the DL course, which should be out early in 2019, will largely be a superset of the current course, but updated for the release version of the fastai library (coming out very soon). Pytorch hasn't changed too much in the interim.
667.
by jph00 8y ago
Yes probably. For instance, there's been interesting recent examples from Pinterest and Instacart (amongst others) who found that DL-based approaches for tabular data reduced the engineering and maintenance headaches due to less need f
668.
by jph00 8y ago
Honestly, either is fine. It depends on your learning style and priorities. If you want to jump quickly to getting state of the art results with rich datasets like images and natural language text, then start with the deep learning course.
669.
by jph00 8y ago
Linear regression is a tricky algorithm to use effectively. Many courses have started with this algorithm, but we think it's a mistake. We cover it later in the course. Decision trees are much easier to use correctly, and can be easily
670.
by jph00 8y ago
The whole course is freely available online: http://course.fast.ai/ml
671.
by jph00 8y ago
Jeremy from fast.ai here - I teach this course. Let me know if you have any questions about it (or else head over to http://forums.fast.ai for the dedicated course forum if it's a more specific technical question).
672.
by jph00 8y ago
I'm often guilty of this too - but we really should put the (tm) there. It's nice that they made code of the algorithm publicly available and all they ask is that we respect their trademark in return. I think that's more than
673.
by jph00 8y ago
Jeremy from fast.ai here - I teach this course. Let me know if you have any questions about it (or else head over to http://forums.fast.ai for the dedicated course forum if it's a more specific technical question).
674.
by jph00 8y ago
OP makes surprising claims about J. Here is some reading which hopefully will help you make up your own mind about whether the claims are accurate. J has a rich range of composition operations (far more than mentioned in OP) which work well
675.
by jph00 8y ago
> Smart Reply, WaveNet, Translate, Speech Recognition, RankBrain, Photo/Image Search, Retinal diagnosis, Waymo (soon), and numerous other components that are used in Android and the Assistant. These aren't examples of just pr
676.
by jph00 8y ago
cuDNN certainly does, if you use the CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD option (and you can ask it to automatically pick the best approach if you like).
677.
by jph00 8y ago
The only person from fast.ai that comments on HN is me. Any other comments that mention fast.ai are from users of the courses or software.
678.
by jph00 8y ago
The ULMFiT, ELMO, and OpenAI Transformer papers are all quite readable and linked from the article. Sebastian and I also wrote an introduction to ULMFiT here: http://nlp.fast.ai/classification/2018/05/15/
679.
by jph00 8y ago
The phrase "Imagenet moment" is generally used to refer to the success of deep learning in the ILSVRC 2012 competition, which used the Imagenet dataset. This is the case in this article.
680.
by jph00 8y ago
Well spotted - this is where I first created the algorithm that became ULMFiT! I wanted to show an example of transfer learning outside of computer vision for the course but couldn't find anything compelling. I was pretty sure a langua
681.
by jph00 8y ago
The naming convention is fully documented and based on decades of research : https://github.com/fastai/fastai/blob/master/docs/style.md It's fine if you don't like it (although it may be y
682.
by jph00 8y ago
Thanks for your work on the reproducibility challenge - we read your results and found them interesting. The good news is that we had a lot more success, although it took many months of work to really make it sing, and bringing in AdamW was
683.
by jph00 8y ago
> It’s a form of credit inflation. Researchers want sweet, sweet deep learning hype credit for fame & fortune. The authors of the AdamW or super-convergence papers did not write the blog post - we did. There is no connection between
684.
by jph00 8y ago
> the published result is almost surely not repeatable in each nuanced case for day to day practitioners that have custom layers, extra regularization, heterogeneous hardware for training As shown in the post and linked code, it works ve
685.
by jph00 8y ago
Jeremy from fast.ai here. Thanks for the thoughtful summary. A few more clarifications which hopefully are of some interest: - Sylvain tried AdamW + 1cycle on many different datasets and with different architectures. The 5-10x improvement o
686.
AdamW and Super-convergence is now the fastest way to train neural nets(fast.ai)
6 points by jph00 8y ago|0 comments
687.
by jph00 8y ago
I tested adding things like tweets and reddit comments to our language model, and it didn't help the target model at all, even when that used less formal language. Note however that the fine tuning stage adapts to the target corpus - i
688.
by jph00 8y ago
This paper is really important: it shows that transfer learning can be applied to a wide variety of NLP problems with great success. They show state of the art results on nearly every major class of NLP problem. The basic approach is the sa
689.
by jph00 8y ago
It would be a great dsl for tensor comprehensions. Also it's a better notation for much (most?) of the math of data science. I wish papers used it. Then the math would be unambiguous and testable and maybe even correct.
690.
by jph00 8y ago
For some reason the field of AI has become an area where famous men who have no expertise in the field feel they have an obligation to tell us all about their opinions on it. Equally oddly, prestigious outlets then decide that these opinion

More ›