Understanding the Representation and Computation of Multilayer Perceptrons: A Case Study in Speech Recognition
By Tasha Nagamine, Nima Mesgarani. 2017
32 citations recorded by the source corpus.
Despite the recent success of deep learning, the nature of the transformations they apply to the input features remains poorly understood. This study provides an empirical framework to study the encoding properties of node activations in various layers of the network, and to construct the exact function applied to each data point in the form of a linear transform. These methods are used to discern and quantify properties of feedforward neural networks trained to map acoustic features to phoneme labels. We show a selective and nonlinear warping of the feature space, achieved by forming prototypical functions to account for the possible variation of each class. This study provides a joint framework where the properties of node activations and the functions implemented by the network can be linked together.
Subfields: Model Architectures, Speech & Audio
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Model Architectures
- Ron J. Weiss
- Aäron van den Oord
- Yi Luo
- Nima Mesgarani
- Samy Bengio
- Thomas Hain
- Jan Chorowski
- Chung-Cheng Chiu
In Speech & Audio
- Nima Mesgarani
- Yi Luo
- Ron J. Weiss
- James R. Glass
- Thomas Hain
- Samy Bengio
- Wei-Ning Hsu
- Shinji Watanabe
Citations in the corpus
- Unsupervised Speech Representation Learning Using WaveNet Autoencoders (2019, background): “speaker gender and identity, from phonetic content, properties which are consistent with internal representations learned by speech recognizers [13], [14].”
- TaSNet: Time-Domain Audio Separation Network for Real-Time, Single-Channel Speech Separation (2017, background): “frequency resolution for accurate speech separation. Further analysis of the network representation and transformation may lead to better understanding of how the network separates competing speakers [28]. (a) (b) Fig. 2. Frequency response of basis signals in (a) causal and (b) noncausal networks. ”
- Unsupervised Adaptation of Acoustic Models for ASR Using Utterance-Level Embeddings from Squeeze and Excitation Networks (2019, background): “It is important to include SE blocks in bottom layers (near to input) since top layers are more invariant to speakers and other acoustic conditions [43].”