Hacker News Reader

Comment by jhartmann | Hacker News Reader

It is always impressive to me how both Professor LeCun, Bengio & Hinton stuck to their guns and worked on these problems while others were not so interested. I'm very glad Hinton's group at DNN Research was able to blow away the competition in the Imagenet challenge. Now many more people are working on these ideas and some really amazing things have been accomplished in a very short time. I can't wait to see where we are in five years, and I love that LeCun's points out the areas where we should focus and what we are not good at yet.

View on HN

Do you have any information about what were the hot topics just before this new resurgence of NNs (meaning around late 90's, and early 2000's).

I know symbolic AI was big in 60s, and 80s, but not sure about recent past.

Probabilistic models. Recent research often focuses on Bayesian models.

Probabilistic models have never really gone away. This presentation by LeCun actually suggests embedding neural networks inside of various types of probabilistic models: factor graphs and conditional random fields. This is, for example, how speech recognition works: the output of a neural network is fed into a probabilistic model (a hidden Markov model).

Actually, state of the art speech recognition has switched over to having a Recursive Neural Network directly run over the audio input. Take a look at the paper at http://arxiv.org/abs/1412.5567 and http://usa.baidu.com/deep-speech-lessons-from-deep-learning/

However combining learning features with other systems is a very powerful approach and combining SVM's on top of the learned features of a Neural Network I would say is common. I personally am more interested in approaches like Deep Fried Convnets (http://arxiv.org/abs/1412.7149) that combine kernel methods as part of the Neural Networks themselves.

Not to nitpick. I just want people to realize there are actually recursive nets that rely on a parser to be built (this is the recursive net that relies on backpropagation through structure). Then there is the recurrent net (LSTMs,multimodal) that rely on backpropagation through time.

Talking to some of the users of Recursive nets, they will be renaming them to tree rnns which should help clear up confusion a bit.

I know that Andrew Ng and colleagues say that they don't use HMMs. I haven't spoken with them (I haven't seen them at speech conferences) so I do not know whether they actually believe this themselves.

I believe the best comparison between "CTC" (which is billed as recurrent neural networks without the HMMs) and the traditional approach is by people at Google, Sak et al, "Learning Acoustic frame labeling for speech recognition with recurrent neural networks", ICASSP 2015. (I can't find a PDF online.)

state of the art is still very much using WFSTs and DNNHMMs. IBM and Google are still beating baidu.

In the case of vision, there were a lot of things going on, but support vector machines, sliding window search, descriptors with invariances to different transforms like SIFT and HOG, spatial pyramids, deformable parts models, and mixtures of gaussians are all hot topics that you'll regularly see in papers from the early 2000's. There was a lot of work on improving runtime for these techniques going on, since training SVM experts and evaluating anything over a sliding window search space is very expensive. You'll also see a lot of work on approaches rooted in graph theory like conditional random fields, different flavors of Markov models, min-cut/max-flow based algorithms, and other types of probabilistic graphical models. Most of this stuff is still in widespread use; AI is a big field.

Support Vector Machines was one of the hot topics.

As far as I remember, everyone was jumping on top of SVMs and Markov Models.

I really like the book "Pattern Recognition and Machine Learning" by Christopher Bishop. It's packed full of all the latest-and-greatest algorithms.

WRT your question, an interesting "feature" of that book is that it was published just before deep neural networks started taking off, so there's no mention of DNNs in the book. You can see what the world was like right when they started taking off.

kernels

Reply on news.ycombinator.com