HN Explorer

Verifying Schmidhuber's Critique of Turing Award for Bengio and Hinton and LeCun

5 points by P-NP ↗ HN
Schmidhuber's critique contains over 200 references http://people.idsia.ch/~juergen/critique-turing-award-bengio-hinton-lecun.html. I have spent more than five hours on verifying them. Now I am kind of an expert on this. It helped that I already knew several of his papers and earlier posts i.e. "Deep Learning: Our Miraculous Year 1990-1991" http://people.idsia.ch/~juergen/deep-learning-miraculous-year-1990-1991.html and "Critique of Honda Prize for Dr. Hinton" http://people.idsia.ch/~juergen/critique-honda-prize-hinton.html

I checked about 95% of Schmidhuber's points of criticism, especially the important ones in Sec. I, II, XIV, XVII, XVIII, XII, XIX, XXI, XVI. They seem fully justified.

Have a look at the astonishing Sec. XVII which lists eight of his direct priority disputes with Bengio and Hinton. And his advice for PhD students in the conclusion.

This is not a glorious chapter in the annals of the Turing award.

7 comments

[ 0.19 ms ] story [ 28.8 ms ] thread
p1esk
Yes, I think he should have taken Bengio’s place.
P-NP
But Hinton also republished lots of stuff first published by Schmidhuber, without citing it. See Sec. XVII (5-8), I, II. Hinton also used methods invented by other authors without citation. Same for LeCun. See Sec. XII, XIX, XXI, XVIII. One could argue that Schmidhuber should have received the full award, without sharing it with anybody, except maybe with his student Hochreiter.
p1esk
It’s not about citing others. It’s about accomplishments. As far as accomplishments Schmidhuber > Bengio.
P-NP
I agree. It’s about accomplishments.

And let's not forget the accomplishments of the other pioneers: Lainnainmaa, the inventor of backpropagation 1970. Ivakhnenko, the inventor of the first deep learning networks that learn internal representations 1965. Fukushima, the inventor of the convolutional neural network architecture 1979. Waibel, who combined backpropagation with convolutions 1987. They are prominently mentioned in the critique in Sec. XII, XIX, XXI, XVIII, XX and others.

p1esk
To me the main accomplishments which led to the beginning of the current AI summer in 2012 are:

1. Convnet (Lecun) 2. LSTM (Schmidhuber) 3. Scaling up the convnet to win ILSVRC (Hinton)

Strictly speaking it was just 1 and 3 which jumpstarted the revolution in computer vision, but the importance of LSTM for sequence modelling (language and speech) is hard to overestimate.

P-NP
1. Schmidhuber's Sec. XVIII points out that the basic convnet architecture with convolutional and downsampling layers is actually due to Fukushima (1979). In 1987, NNs with convolutions were combined by Waibel with weight sharing and backpropagation, before LeCun.

3. Sec. D and XIV say that the convnet was scaled up already in 2011 through the DanNet from Schmidhuber's lab. DanNet was the first convnet to win a computer vision contest. It also was the first convnet to achieve superhuman performance and had a "monopoly on winning CV competitions." It won 4 CV contests before Hinton's AlexNet won ILSVRC.

It's true. I checked the references.