Information Technology Reference
In-Depth Information
Figure 8. Artificial neural network layer using the fanout composite
horizontal forking. For example the output of
the
Spectrum
calculation is used as input to the
Centroid
,
Rolloff
and
Flux
MarSystems
. The input
to texture memory is 4 observations (the features)
by 1 sample and the output is 4 observations by
40 samples consisting of the last 40 feature vec-
tors (approximately corresponding to 1 second).
Means and variances of the feature vectors over
the texture window are calculated and the input
to the classifier is an 8-dimensional feature vector
(4 means and 4 vairances). The entire network can
be created at run-time without requiring any code
recompilation. The complete feature extraction
front-end described in Tzanetakis and Cook (2002)
has been implemented as a dataflow network in
MARSYAS
in a similar fashion.
vations) based on the
Fanout
semantics become
the input to each individual neuron (Ni)
i
) of the
layer. Each neuron forms a weighted sum (with
weights specific to each neuron) of the input, ap-
plies a sigmoid function to the sum and outputs
a single output. The outputs using the
Fanout
semantics are stacked as observations y
1
, y
2
, y
3
(one for each neuron) ready for processing for
the next layer. Figure 8 illustrates this process
graphically (left side) and contrasts it with Explicit
Patching (right side). In
MARSYAS,
creating an
ANN using an
AnnNode MarSystem
is simply
a
Series
of
Fanout
of
AnnNodes
. More specifi-
cally seriesNet(fanoutLayer1, fanoutLayer2, …,
fanoutLayerM) where fanoutLayer1(annNode11,
annNode12, …,annNode1N). All the connections
are created implicitly.
Figure 9 shows a dataflow network for extract-
ing audio features for real-time music/speech
classification.
Series
connections are top to
bottom and
Fanout
connections are shown by
Flexible Scheduling
Scheduling is central to any computer music
system. A scheduling request consists of an event
Search WWH ::
Custom Search