Interpretation of Neural Networks and Anatomy of Speech Recognizer

  Machine learning

Distillation of the neural network into a soft decision tree - Lukáš Marták and Pavel Konečný
Neural networks are frequently compared to a black box. It is difficult to interpret the decision process. What to do if a customer ask for it? We did implement a method, which does emulate the decisioning of the neural network by a soft decision tree. Learn more about our Keras application of the scientific paper published by Nicholas Frosst and Geoffrey Hinton (Submitted on 27 Nov 2017). We will share the experiences and results of the on a client sensor dataset (audio data) and MNIST.

Anatomy of a Speech Recognizer - Karel Veselý
Speech recognition is a relatively easy way of interacting with various devices from the end-user perspective. However, not so many people have an idea about its inner workings. In this talk, I’d like to make a ‘gentle introduction’ providing a picture of the ‘anatomy of a speech recognizer’, and it’s possible variations which were created by the community of researchers. I’ll definitely touch the related topics like signal processing, machine learning and hypothesis search. And I’d also like to introduce the Semi-supervised training, which is a paradigm of improving an existing system by retraining it with untranscribed data.

Language: English

Pavel Konečný a Lukáš Marták
Pavel Konečný, CEO and Co-Founder of Neuron soundware. Pavel studied. Realized by Lukáš Marták.

Karel Veselý
Karel Veselý is a speech recognition in-sider for nearly a decade. After joining the Speech@FIT lab in 2007 he went for ‘speech technology’ oriented Erasmus to IUP/IUT Avignon. Then, back in Brno, he implemented a diploma project TNet (a CUDA trainer of neural networks). Later, as a young PhD student (in 2010), he joined the core team of developers of the Kaldi project, and co-organized the 4 Kaldi workshops in Brno (2010, 2011, 2012, 2013). Kaldi is an extremely popular open-source platform for developing speech recognition systems, it is used both in academia and industry. And, Karel contributed the neural network training codebase `nnet1’, which was for a while giving the best results. With his many interests, he finally converged to focus on semi-supervised training of neural network based acoustic models, which is the topic of his thesis (defended in spring 2018). At present he works both for Brno University of Technology and a start-up SoapBox Labs.

- 17:45 - 18:00 - Your arrival
- 18:00 - 18:40 - Distillation of the neural network into a soft decision tree - Lukáš Marták and Pavel Konečný
- 18:40 - 18:50 - Short break
- 18:50 - 19:30 - Anatomy of a Speech Recognizer - Karel Veselý
- 19:30 - 22:00 - Networking in Bitcoin Coffee