Hacker News Reader

Comment by platz | Hacker News Reader

how relevant are hidden markov models now? Do they still have unique strengths that aren't covered by other, newer models?

View on HN

Some state-of-the-art industrial speech recognition [0] is transitioning from HMM-DNN systems to "CTC" (connectionist temporal classification), i.e., basically LSTMs. Kaldi is working on "nnet3" which moves to CTC, as well.

Speech was one of the places where HMMs were _huge_, so that's kind of a big deal.

[0] http://googleresearch.blogspot.com/2015/09/google-voice-sear...

I'm kind of more interested in if HMM's are more viable as a general-purpose tools rather than their applicability in cutting-edge research; or if they should generally be avoided for common tasks unless the domain is simple enough.

I wonder how much of ANN popularity is due to the inherent strength of the model, and how much of it is due to large corporations funding/hiring people who happen to be interested in that particular flavor of machine learning.

HMMs are only a small subset of generative models that offers quite little expressiveness in exchange for efficient learning and inference.

Thanks

IMO, anything that be done with an HMM can now be done with an RNN. The only advantage that an HMM might have is that training it might be faster using cheaper computational resources. But if you have the $$$ to get yourself a GPU or two, this computational advantage disappears for HMMs.

It's almost sad – I always felt HMMs had a unique elegance that is lost on the brute-force NN approach.

HMMs also used to produce more meaningful intermediate results, although I guess these 'DeepDream' images set a new record in that regard :)

Not really. The first order Markov property is really naive when compared to an LSTM.

Reply on news.ycombinator.com