Beruflich Dokumente
Kultur Dokumente
:Written By
Gassan Tabajah
Ron Adar
Gil Rapaport
Abstract
The Liquid State Machine (LSM) is a recently
developed computational model with interesting
properties.
It can be used for pattern classification, function
approximation and other complex tasks.
Contrary to most common computational models, the
LSM does not require information to be stored in
some stable state of the system:
The inherent dynamics of the system are used by a
memoryless readout function to compute the output.
Abstract
In this paper we present a case study of the performance of the
Liquid State Machine based on a recurrent spiking neural
network by applying it to a well known and well studied
problem:
Speech recognition of isolated digits.
We evaluate different ways of coding the speech into spike
trains.
In its optimal configuration, the performance of the LSM
approximates that of a state-of-the-art recognition system.
Another interesting conclusion is the fact that the biologically
most realistic encoding performs far better than more
conventional methods.
Introduction .1
Many complex computational problems have
a strong temporal aspect:
Not only the value of the inputs is important, but
also their specific sequence and precise
occurrence in time.
Tasks such as speech recognition, object
tracking, robot control or biometrics are
inherently temporal, as are many of the tasks
that are usually viewed as ‘requiring
intelligence’.
Introduction .1
However, most computational models do not
explicitly take the temporal aspect of the input into
account or transform the time-dependent inputs to
static input using.
e.g., a tapped delay line.
These methods disregard the temporal information
contained in the inputs in two ways:
The time-dependence of the inputs within a certain time
window is compressed into a static snapshot and is
therefore partially lost.
The temporal correlation between different windows is not
preserved.
Liquid State Machine (LSM)
The Liquid State Machine (LSM) avoids these problems by
construction.
The LSM is a computational concept (its structure is depicted in
Fig. 1):
Liquid State Machine (LSM)
A reservoir of recurrently interacting
nodes is stimulated by the input u(t),
A running or liquid state x(t) is
extracted and a readout function fM
converts the high-dimensional liquid
state x(t) into the desired output y(t)
for the given task.
Liquid State Machine (LSM)
The loops and loops within-loops which occur
in the recurrent connections between the
nodes in the reservoir cause a short-term
memory effect to occur:
The influence of inputs fed into the network
‘resonate’ for a while before dying out.
Hence the name liquid: ripples in a pond also
show the influence of past inputs.
This property is called temporal integration.
Types of liquids
The reservoir or liquid can be any type of
network that has sufficient internal dynamics.
Currently several types of liquids have been
used:
The echo state machine using a recurrent analog
neural network.
A real liquid (water) in a bucket.
A delayed threshold logic network.
A Spiking Neural Network (SNN).
The type of liquid we will be using in this article.
Readout function
The actual identity of the readout function is also not explicitly
specified;
Indeed, it can be any method of statistical analysis or pattern
recognition.
Possible readout functions include linear projection, Fisher
discriminant, a perceptron, a feed forward MLP trained with
backpropagation or even a Support Vector Machine.
Note that the liquid itself is not trained, but chosen a priori:
A heuristic is used to construct ‘interesting’ liquids in a random
manner.
Only the readout function is adapted so that the LSM performs
the required task.
Separation between the
liquid and its readout
The separation between the liquid and its readout
function offers two considerable advantages over
traditional Recurrent Neural Networks (RNN).
First, the readout function is generally far easier to train
than the liquid itself, which is a recurrently connected
network.
Furthermore, the structure of the LSM permits several
different readout functions to be used with the same liquid,
which can each be trained to perform different tasks on the
same inputs—and the same liquid.
This means that the liquid only needs to be computed once,
which gives the LSM an inherent parallel processing capability
without requiring much additional processing power.
Similarity between LSM and SVM
The LSM is a computational model that bears strong
resemblance to another well-known paradigm:
That of the kernel methods and the Support Vector Machine (SVM).
Here, inputs are classified by first using a non-linear kernel function
to project the inputs into a very high-dimensional (often even
infinitely dimensional) feature space, where the separation
between the classes is easier to compute (and can often even be
linearly separable).
The LSM can be viewed as a variation:
Here too, the inputs are non-linearly projected into a high-
dimensional space (the space of the liquid state) before the
readout function computes the output.
The main difference is the fact that the LSM—contrary to SVMs
—has an inherent temporal nature.
Similarity between LSM and RNN
Another similarity can be found in a the state-
of-the-art RNN training algorithm called
Backpropagation Decorrelation presented in
[6], which was derived mathematically.
It has been shown that training a RNN with
this algorithm implicitly also keeps a pool of
connections almost constant and only trains
the connections to the output nodes.
This behavior is very similar to the principle of the
LSM, specifically when using a perceptron as
readout function.
Filter approximation
In [7] it has been shown that any time
invariant filter with fading memory can
be approximated with arbitrary
precision by a LSM, under very
unrestrictive conditions.
In practice, every filter that is relevant
from a biological or engineering point of
view can be approximated.
Article structure
This article is structured as follows:
In Section 2 we first detail the setup used to perform the
experiments.
In Section 3 we briefly introduce and subsequently test three
different speech front ends which are used to transform the
speech into spike trains.
The effects of different types of noise that are commonly found
in real world applications are tested in Section 4.
In Section 5 we make some relevant comparisons with related
isolated word recognition systems.
In Section 6 we draw some conclusions about the speech
recognition capabilities of this computational model.
In Section 7 we point out some possible opportunities for
further research.
Experimental Setup
Matlab LSM toolbox
Liquid is a
recurrent network (contains loops)
of leaky integrate and fire neurons
arranged in a 3D column (pile)
Parameters taken from original LSM
paper
Network structure
Auditory
nerve
Cochlea
The LSM was compared to the best results from one of the
books in the references, which was the “Log Auditory Model”
(LAM) with noise robust front-end.
’Input with noise cont
The paper have looked the SNN interpretation of the LSM to the
task of isolated word recognition with a limited vocabulary.
The results showed that LSM is well suited for the task, and that
the performance is far better using biological model (LPEM).
Find out what causes the big difference between LPE model to
the traditional MFCC methods.