Unsupervised Neural Coding of Nightingale Songs Using Deep Autoencoders
-
1
FU Berlin, Germany
-
2
Uni Freiburg, Germany
In this work we tested different deep autoencoder networks for the unsupervised feature extraction of nightingale song spectra. Nightingales are versatile singers. Some individuals can sing up to 600 different songs. Biologists analyzing nightingale communication have to classify audio recordings by comparing hundreds of unknown songs to a database. This work is done mostly manually by matching spectrograms visually. An automatic or semi-automatic method for song classification would speed up this tedious process. We have utilized a recently published learning method to train multi-layered (‘deep’) artificial neural networks to reduce the dimensionality of – and find correlations within the spectrogram data in an unsupervised manner. We propose several preprocessing steps and network topologies to find low dimensional representations of nightingale songs. First, the audio data is filtered with a band pass to reduce low-frequency noise, e.g. of nearby cars. Then, we normalize the volume and down-scale the spectrograms to 256 x 400 points. This matrix is used as the input layer to the network. The next layer extracts visual features like edges and corner points. Each neuron in that layer serves as a feature detector and shares its incoming weights with different ‘receptive fields’ in the input layer and thus establishes repetition- and shift invariance. The output of this layer will be fed to the next three layers that serve for the dimensionality reduction and are trained as proposed by Hinton (2006). The weights of the network are tuned by comparing the input of the network to its reconstruction: By feeding an input song to the network, a specific code can be read from the last, 16-dimensional, code layer. By projecting back the activity of this layer to the receptive field, using the same weights, it is possible to reconstruct its original excitation; a procedure we use also to measure the quality of the code. Once the training is complete it is possible to classify unknown songs using the low dimensional code with an additional classification layer or other standard classification methods.
Keywords:
data analysis,
deep autoencoder,
machine learning,
nightingale songs,
unsupervised learning
Conference:
BC11 : Computational Neuroscience & Neurotechnology Bernstein Conference & Neurex Annual Meeting 2011, Freiburg, Germany, 4 Oct - 6 Oct, 2011.
Presentation Type:
Poster
Topic:
data analysis and machine learning (please use "data analysis and machine learning" as keyword)
Citation:
Landgraf
T,
Brachmann
A,
Blum
M and
Rojas
R
(2011). Unsupervised Neural Coding of Nightingale Songs Using Deep Autoencoders.
Front. Comput. Neurosci.
Conference Abstract:
BC11 : Computational Neuroscience & Neurotechnology Bernstein Conference & Neurex Annual Meeting 2011.
doi: 10.3389/conf.fncom.2011.53.00078
Copyright:
The abstracts in this collection have not been subject to any Frontiers peer review or checks, and are not endorsed by Frontiers.
They are made available through the Frontiers publishing platform as a service to conference organizers and presenters.
The copyright in the individual abstracts is owned by the author of each abstract or his/her employer unless otherwise stated.
Each abstract, as well as the collection of abstracts, are published under a Creative Commons CC-BY 4.0 (attribution) licence (https://creativecommons.org/licenses/by/4.0/) and may thus be reproduced, translated, adapted and be the subject of derivative works provided the authors and Frontiers are attributed.
For Frontiers’ terms and conditions please see https://www.frontiersin.org/legal/terms-and-conditions.
Received:
23 Aug 2011;
Published Online:
04 Oct 2011.
*
Correspondence:
Mr. Tim Landgraf, FU Berlin, Berlin, Germany, tim.landgraf@fu-berlin.de