Reconstructing Speech Input from Convolutional Neural Network Activity
Keywords
Loading...
Authors
Issue Date
2015-07-13
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Convolutional Neural Networks (CNNs) applied to the auditory domain have achieved great results.
However, little research has been performed to uncover the underlying mechanisms that allow auditory
CNNs to reach these achievements. This exploratory research attempts to help uncover these mechanisms
by using a CNN's activation patterns generated by speech inputs to reconstruct those inputs. It is found
that training a decoder to decode activity patterns to a preliminary reconstruction and consequently fine-tuning that reconstruction through further back propagation generates the best results. The reconstructions
show that the network preserves a good representation of the input up to and including the fully connected
units. Reinforcingly, this representation appears to be suited to speech, instead of, for example, audio in
general. Furthermore, it becomes apparent that the network is insensitive to input intensity as well as to
the input's activation scale on the time-domain. Further research is required to discover more properties
of CNNs applied to the auditory domain.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen