Recently, deep learning has been successfully introduced to the field
of music information retrieval. Several deep learning models have been
applied to different MIR tasks, such as deep neural networks,
convolutional neural networks, and deep belief networks. As an
alternative, we recently proposed the compositional hierarchical model
(CHM). To overcome some of the limitations by other deep approaches
(need for large datasets, black-box approaches), the CHM proposes a
transparent structure in a form of a generative model. Its main
features are unsupervised learning of a hierarchical representation of
input data, transparency, which allows for insights into the learned
representation, the ability to extract knowledge from a small dataset,
and robustness and speed.
The model consists of multiple layers, each composed of a number of
parts, unsupervisedly learned from the music input. Parts in each layer
are compositions of parts from previous layers based on statistical
co-occurrences as the driving force of the learning process.
In this talk, we will present how the CHM has been applied to three
different music information retrieval tasks: polyphonic music
transcription, chord estimation and melodic pattern discovery.
We will also show possible applications of the CHM to other tasks and
domains and elaborate on connecting the CHM to other ML approaches.
Time: Wednesday, 18th January 2017, 4:00 p.m. sharp
Location: Oesterreichisches Forschungsinstitut
fuer Artificial Intelligence, OFAI
Freyung 6, Stiege 6, 1010 Wien
FUER ARTIFICIAL INTELLIGENCE
Univ.-Prof. Ing. Dr. Robert Trappl