Hacker News Reader

Comment by nabla9 | Hacker News Reader

Schmidhuber's significance in the field is not at the same level. He invented one type of neural network, but the "Canadian mafia" launched the whole Deep Learning revolution with multiple breakthroughs and theoretical understanding.

LSTM's are clever and they were bleeding edge up to 2014, but once they were understood better as a bypass mechanism, attention, context vectors and averaging networks and causal convolution are starting to replace them.

View on HN

Ironically Schmidhuber's lab also came up with a more general version of bypassing http://arxiv.org/abs/1505.00387 to train deep networks (with similar performance), of which the popular ResNet is a special case.

But highway network doesn't work as well as ResNet, and ResNet is simpler.

Highway connections are actually used heavily in NLP in combination with LSTMs.

And replaced by residual connections in transformers, which are absolutely dominating LSTMs now.

Transformer-XL uses recurrence, and most NLP SOTA is still with LSTMs. I’m not sure I’d expect attention mechanisms to fully replace recurrence.

Highway networks with scalar gating beats ResNet by a small margin.

Putting aside the whole Schmidhuber debate - where are people getting this idea that causal convolutions are anywhere near the prominence of RNNS/LSTMs?

As far as I'm aware, causal convolutions were used in WaveNet (and subsequent models) and a small number of NLP applications. Meanwhile, LSTM-based models are used in just about every NLP paper, and at least a baseline in the newer ones more dominated by Transformers.

I think LSTMs have been mostly abandoned. I don't see much new work using them, vs CNNs or transformers.

Aren't causal convolutions basically CNNs with masking?

"Schmidhuber's significance in the field is not at the same level"

Out of curiosity, at the same level as any of the other three? I could see it from Hinton, maybe LeCunn, but Bengio?

So ... Schmidhuber's was the seminal work and the later ones were incremental improvements that came from an understanding of Schmidhuber's work ?

That's very ignorant to say that attention or convolution is anywhere close to the expressivity of RNNs(LSTM). Instead of reading a few cherry-picked results from a model with extremely tuned hyperparameters, pick any random set of tasks and experiment yourself. Based on my experience, LSTM always performs better than all unless you brutely search over hundreds of hyperparameters configs.

> That's very ignorant to say that attention or convolution is anywhere close to the expressivity of RNNs(LSTM).

But that's exactly the point of the Transformer model, with a paper aptly titled "Attention is all you need" [1]. And the Bert architecture, based in this idea, seems to be doing well. And they claim to be bery flexible, too[2].

Maybe that's what you meant with "unless you brutely search over hundreds of hyperparameters configs", but then again, isn't that what NNs are about anyway?

[1] https://arxiv.org/abs/1706.03762

[2] https://arxiv.org/abs/1810.04805

The success of Transformers aside, I'm not sure you should be relying on model titles for anything, lest we forget papers like "One Model To Learn Them All" [1].

[1] https://arxiv.org/abs/1706.05137

Schmidhuber did way more than LSTMs. And they still are SOTA on many tasks. You clearly are not very familiar with the field.

That third sentence is not necessary.

I'm not saying that Schmidhuber is not important pioneer in the field.

When you put researchers or groups into order of importance, the work from this trio comes before others, including Schmidhuber. The Turing Award is given to those at the top and it's not inclusive.

If I could give Turing Award for someone in the field who has not received it yet, I would give it to Vladimir Vapnik.

Sorry.

>If I could give Turing Award for someone in the field who has not received it yet, I would give it to Vladimir Vapnik.

So would I, but I don't think that will happen. He's is pretty much an antithesis to everything Deep Learning is about. Theory-first over trial-and-error, math over intuition, small datasets over big data, advances in understanding vs advances in results.

I haven't seen a single discussion of his paper on combining classifiers[1] anywhere on the web.

[1] - http://jmlr.csail.mit.edu/papers/volume17/16-137/16-137.pdf

Whoa. Personal attacks will get you banned here. We've warned you about this more than once before. Please review https://news.ycombinator.com/newsguidelines.html and follow the rules when posting here. When you do, don't miss this one, since it should have changed your assumption about the GP:

"Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize. Assume good faith."

You're right. Sorry.

Just a note, I did not find it any more rude than other comments that get a free pass. Commenting that more familiarity with the literature is necessary can be a very factual claim.

It was an obvious violation of HN's rules. If you see other comments that violate the rules going unmoderated, the likeliest explanation is that we didn't see them; we don't come close to seeing everything that gets posted. In that case the thing to do is to flag the comment (described at https://news.ycombinator.com/newsfaq.html). In egregious cases, you can email hn@ycombinator.com as well.

Reply on news.ycombinator.com