eugenhotaj

9 ms·

31.
Writing Web Servers from First Principles(medium.com)
3 points by eugenhotaj 6y ago|0 comments
32.
by eugenhotaj 7y ago
I disagree, PAC bounds are usually way too loose to be useful in practice when it comes to NNs. I think part of the reason is that the way we measure model complexity for NNs does not correspond very well with their ability to generalize, a
33.
Necessary Tricks for High Quality Neural Style Transfer(towardsdatascience.com)
1 points by eugenhotaj 7y ago|0 comments
34.
by eugenhotaj 7y ago
Misleading -- you can just tuck in your fu manchu
35.
by eugenhotaj 7y ago
> I'm particularly proud of this meta approach and I am actually thinking this could become huge: the same thing can be done for hyperparameter optimization in machine learning tasks. There is already a substantial field of Machine
36.
Is It Feasible to Expect Machine Learning to Work?(towardsdatascience.com)
2 points by eugenhotaj 7y ago|0 comments
37.
by eugenhotaj 7y ago
I would expect it to (but I haven't thought about it too deeply so I could be extremely wrong). My thinking is as follows: At the end of the day, all we're doing is maximum likelihood estimation. So we're trying to find model
38.
by eugenhotaj 7y ago
I've seen some research [1] where the authors use beam search with an explicit diversity penalty to get around the repetition problem. They seem to get good results. [1] https://arxiv.org/pdf/1610.02424.pdf
39.
by eugenhotaj 7y ago
Hi, author here. Yea I did the delimiting you mentioned when "training" a bigram model. For GPT-2 I was mostly interested in how well the model would be able to pick up signals from the raw data so I didn't do any kind of pre
40.
by eugenhotaj 7y ago
Really cool stuff, thanks for sharing!
41.
Finetuning GPT-2 to Generate Beatles Lyrics(towardsdatascience.com)
55 points by eugenhotaj 7y ago|14 comments
42.
Deep Learning: Our Miraculous Year 1990-1991(people.idsia.ch)
156 points by eugenhotaj 7y ago|34 comments
43.
by eugenhotaj 7y ago
I have not seen the presentation yet. Still, I am willing to bet anyone a non-trivial amount of money that we do not see people flying next year.
44.
by eugenhotaj 7y ago
Maybe I'm misunderstanding numbers, but if the first attempt was aborted then this is the second attempt?
45.
Generating Beatles' Lyrics with Machine Learning(towardsdatascience.com)
1 points by eugenhotaj 7y ago|0 comments
46.
by eugenhotaj 7y ago
> or great tools like OP's for setting up your own I don't see how I was trashing OPs hard work. I'm simply pointing out that there's more that Medium provides than just a way to host a blog. No need to read so much n
47.
by eugenhotaj 7y ago
I honestly don't get posts like these. IMHO the main selling point of Medium is not that you can easily set up a blog, there's plenty of platforms that let you do that easily, or great tools like OP's for setting up your own.
48.
Ask HN: Does Anyone Use Code Folding?
2 points by eugenhotaj 7y ago|6 comments
49.
Making Rockets Hover with PID Controllers(medium.com)
2 points by eugenhotaj 7y ago|0 comments
50.
by eugenhotaj 8y ago
Hi, author here. Thanks for reading! Yea these are all great points and most are brought up in the article.
51.
by eugenhotaj 8y ago
This is equivalent to claiming that if you randomly sample from the real number line the probability of sampling a natural number is 0, therefore natural numbers can't exist. (Of course I doubt the author(s) were making these claims in
52.
by eugenhotaj 8y ago
Agreed, the code is not meant to be a good solution. This was actually my starting point for an implementation of an "evolution" based algorithm that I wanted to experiment with. I was surprised that completely random models were
53.
Show HN: OpenAI's Cartpole Can Be Optimally Solved in ~10 Random Initializations(github.com)
1 points by eugenhotaj 8y ago|2 comments
54.
Simple CNN on Cifar10 Using Keras(github.com)
1 points by eugenhotaj 9y ago|0 comments
55.
by eugenhotaj 9y ago
Yea not directly supported by Google, but being built by Google engineer(s).
56.
Neural Network Addition for numbers(github.com)
3 points by eugenhotaj 9y ago|0 comments