Hello everybody
I am planning to give a series of 3 talks at CRIM on recent developments in Boltzmann mahcines, deep belief nets and deep neural nets with applications to speech processing. The target audience is primarily researchers and PhD students working in speech recognition or speaker recognition.
I aim to explain
(i) how to build the deep neural networks for speech recognition that enabled Microsoft to achieve the results reported in the
2011 Interspeech paper by Seide and colleagues
(ii) how Boltzmann machines might be used to replace the generative models currently used in speaker recognition (the universal background model, the ivector extractor, probabilistic linear discriminant analysis and the cosine distance metric)
I will assume that the audience is familiar with basic probability (I will begin by reviewing variational Bayes and expectation maximization) but not necessarily the machine learning literature.
I will distribute notes on these topics:
(i) Boltzmann machines, Salakhuditnov's algorithms
(ii) Sigmoid belief networks
(iii) Hinton's construction of deep belief nets, backpropagation
Tentatively, I propose giving these talks at 4 p.m. on Monday, Wedensday and Friday, February 6, 8 and 10.
If any of your students are interested please have them get in touch with me so that we can sort out any scheduling conflicts and I can confirm the final arrangements.
cheers
Patrick