arXiv:1502.02444vl [cs.NE] 9 Feb 2015
On the Dynamics of a Recurrent Hopfield Network
Rama Garimella
Department of Signal Processing
Tampere University of Technology
Tampere, Finland
Email: rammurthy@iiit.ac.in
Berkay Kicanaoglu
Department of Signal Processing
Tampere University of Technology
Tampere, Finland
Email: berkay .kicanaoglu @ student.tut.fi
Moncef Gabbouj
Department of Signal Processing
Tampere University of Technology
Tampere, Finland
Email: moncef.gabbouj@tut.fi
Abstract —In this research paper novel real/complex valued
recurrent Hopfleld Neural Network (RHNN) Is proposed. The
method of synthesizing the energy landscape of such a network
and the experimental investigation of dynamics of Recurrent
Hopfleld Network is discussed. Parallel modes of operation (other
than fully parallel mode) in layered RHNN is proposed. Also,
certain potential applications are proposed.
I. Introduction
In the research field of electrical engineering, following
the principle of parsimony, many natural as well as artificial
systems were modeled as linear systems. However it was
realized that many natural/artificial phenomena are inherently
non-linear. For instance, many models of neural circuits re¬
quire non-linear difference/differential equations for modeling
their dynamics. Hopfield effectively succeeded in arriving at
a model of associative memory based on McCulloch-Pitts
neuron. The discrete time Hopfield network represents a non¬
linear dynamical system.
Hopfield Neural Network spurred lot of research efforts on
modeling associative memories (such as the bi-directional as¬
sociative memory). However, it was realized that the Hopfield
neural network is based on undirected graph(for topological
representation of network architecture) and hence does not
have any directed cycles in the network architecture. Thus, in
that sense Hopfield Neural Network (HNN) does not constitute
a recurrent neural network. Many recurrent networks, such
as Jordan-Elman networks were proposed and studied exten¬
sively. After literature survey, we were motivated to propose
an Artificial Neural Network (ANN), in the spirit of HNN,
but the network architecture is based on a directed graph with
directed cycles in it.
Our research efforts are summarized in this paper. In
Section-2, a novel real as well as complex Recurrent Hopfield
Network is discussed. In Section-3, experimental investiga¬
tions of Recurrent Hopfield neural Network are discussed. In
Section-4, some applications are proposed. The research paper
concludes in Section-5.
II. Recurrent Hopfield Network(RHN)
Recurrent Hopfield Neural Network (RHNN) is an Artificial
Neural Network model. It is a nonlinear dynamical system
represented by a weighted, directed graph. The nodes of
the graph represent artificial neurons and the edge weights
correspond to synaptic weights. At each neuron/node, there
is a threshold value. Thus, in summary, a Recurrent Hopfield
neural network can be represented by a synaptic weight matrix,
M that is not necessarily symmetric and a threshold vector,
T. The order of the network corresponds to the number of
neurons. Such a neural network potentially has some directed
cycles in the associated graph.
Every neuron is in one of the two possible states H-1 or
-1. Thus, the state space of Nth order Recurrent Hopfield
Neural Network (RHNN) is the symmetric N-dimensional unit
hypercube. Let the state of /th neuron at time t be denoted by
Vi[t) where Vi{t) G {-fl or — 1}.
Thus, the state of the non-linear dynamical system is
represented by the N x 1 vector, V{t). The state updation
at the ith node is governed by the following equation
N
V,{t + 1) = SigniY, - T,) (1)
where Sign (.) is the Signum function. In other words, Vi(t-\-
1) is H-1 if the term in the brackets is non-negative, otherwise
is -1. Depending on the set of nodes at which the state
updation by Eq. [T] is performed at any time f, the Recurrent
Hopfield Neural Network (RHNN) operation is classified into
the following modes.
■ Serial Mode: The state updation in Eq. [T] is performed
exactly at one of the nodes/neurons at time t.
• Fully Parallel Mode: The state updation in Eq. [T] is
performed simultaneously at all the N nodes/neurons at
time t.
Remark (1). If the matrix M is symmetric, then the dynamics
of Recurrent Hopfield Neural Network (RHNN) is exactly same
as the same as that of Ordinary Hopfield Neural Network
(OHNN).
For the sake of completeness, we now include the dynamics
of Ordinary Hopfield Neural Network (OHNN) with M being
a symmetric matrix [ 1 ]. In the state space of Ordinary Hopfield
Neural Network (OHNN), a non-linear dynamical system,
there are certain distinguished states, called the stable states.
Definition II.l. A state, V{t) is called a stable state if and
only if
V{t) = Sign{MV{t) - T)
( 2 )
Thus, once the Hopfield Neural Network (HNN) reaches
the stable state, irrespective of the mode of operation of the
network, the HNN will remain in that state for ever. Thus, there
is no further change of the state of HNN once a stable state is
reached. The following convergence Theorem summarizes the
dynamics of Ordinary Hopfield Neural Network (OHNN). It
characterizes the operation of neural network as an associative
memory.
Theorem II.l. Let the pair N = (M, T) specify a Ordinary
Hopfield Neural Network (with M being symmetric).Then the
following hold true:
[1 ] Hopfield: If N is operating in a serial mode and the
elements of the diagonal of M are non-negative, the network
will always converge to a stable state (i.e. there are no cycles
in the state space).
[2] Goles: If N is operating in the fully parallel mode, the
network will always converge to a stable state or to a cycle of
length 2 (i.e the cycles in the state space are of length <2).
Remark (2). It should be noted that in [4], it is shown that
the synaptic weight matrix of OHNN can be assumed to be an
arbitrary symmetric matrix.
A. Layered Recurrent Hopfield Neural Network
We now consider a Recurrent Hopfield Neural Network
where the neurons (artificial) are organized into finitely many
layers (with > 1 neurons in each layer). Thus the synaptic
weight matrix of such a neural network can be captured
by a block matrix, W which is not necessarily symmetric.
Traditionally, the convergence Theorem associated with Ordi¬
nary Hopfield Neural Network (OHNN) (i.e. Theorem nm
effectively considered only (i) Serial Mode and (ii) Fully
Parallel Mode. But the arrangement of neurons in multiple
layers naturally leads to operation of the Hopfield network
(Ordinary as well as Recurrent) in other parallel modes of
operation. This is accomplished in the following manner:
• It should be noted that the state vector of the Ordi¬
nary/Recurrent Hopfield Neural Network at time t can
be partitioned into sub-vectors in the following manner.
V(t) =_[V‘^^\t) : 1/(2) (i) .
where l/(^( (t) denotes the state of neurons in layer 1 and
L is the number of layers.
For the sake of notational convenience, suppose that
all the ’L’ layers have the same number of neurons in
it and state updation at any time instant takes place
simultaneously at all nodes in any one layer. This clearly
corresponds to parallel mode of operation which is NOT
the FULLY PARALLEL MODE (if there is a single
neuron in each layer, it corresponds to the serial mode of
operation). The state updation in such a ’parallel mode’
can be captured through the following equation: For
1 < j < L, we have that
L
U«(f + 1) = StgniY, 1U(,-fe)t/W(f) - T,)
fc=i
where ll/(j is the sub-matrix of W and Tj is the
corresponding threshold vector. It can be shown that there
is no loss of generality in assuming that the threshold
vector is a zero vector.
Note: To the best of our knowledge, the dynamics of
even OHNN in other parallel modes of operation (not
fully parallel) has not been fully investigated.
Remark (3). This idea of layering the neurons enables
one to study the dynamics of Recurrent Hopfield Neural
Network with various types of directed cycles in it.
Theoretical efforts to capture the dynamics of such neural
networks are underway. In the Section-3, we summarize
our experimental efforts.
• Energy Landscape Synthesis:
Now we investigate the energy landscape visited by a
Recurrent Hopfield Neural Network (RHNN) and relate it
to the energy landscape visited by the associated Ordinary
Hopfield Neural Network (OHNN).
• Consider the quadratic energy function associated with
the dynamics of a RHNN whose synaptic weight matrix
is the non-symmetric matrix, W. It is easy to see that
V'^(n)WV(n) = V^{n)
= V'^in)WV(n)
where W is the symmetric matrix associated with W, i.e.
symmetric part.
• Now, from the above equation, it is clear that the energy
landscape visited by the RHNN with synaptic weight
matrix W is exactly same as the energy landscape visited
by the OHNN with the symmetric synaptic weight matrix
W. Now, we synthesize the energy landscape of RHNN
(or associated OHNN) in the following manner:
We need the following definition in the succeeding dis¬
cussion.
Definition II.2. The local minimum vectors of the energy
function associated with W are called anti-stable states. Thus,
if u is an anti-stable state of matrix W, then it satisfies the
condition
u = —Sign(Wu).
Using the definitions, we have the following general result.
Lemma 11.2. If a comer of unit hypercube is an eigenvector
of W corresponding to positive/negative eigenvalue, then it is
also a stable/anti-stable state.
Proof: Follows from the utilization of definitions of
eigenvectors, stable/anti-stable states. Q.E.D.
• Synthesis of Symmetric Synaptic Weight Matrix with
desired Energy Landscape:
Let {^} where j G be desired positive
eigenvalues (with pj being the desired stable value) and
let {Xj}{j G {0,1,5'}) be the desired stable states of
W. Then it is easy to see that the following symmetric
matrix constitutes the desired synaptic weight matrix of
Ordinary Hopfield Neural Network (Using the spectral
representation of symmetric matrix W):
s
f=l
It is clear that the synaptic weight matrix is positive
definite. In the same spirit as above, we now synthesize
a synaptic weight matrix, with desired stable/anti- stable
values and the corresponding stable / anti-stable states.
Let { Xj } where j G {0,1,...,5'} be desired orthogo¬
nal stable states and {Yj} where j G {0,1,...,L} be
the desired orthogonal anti-stable states. Let the desired
stable states be eigenvectors corresponding to positive
eigenvalues and let the desired anti-stable states be
eigenvectors corresponding to negative eigenvalues. The
spectral representation of desired synaptic weight matrix
is given by
i=i i=i
where /i^’s are desired positive eigenvalues and —/3j’s
are desired negative eigenvalues. Hence the above con¬
struction provides a method of arriving at desired energy
landscape (with orthogonal stable/anti-stable states and
the corresponding positive/negative energy values).
■
Remark (4). It can be easily shown that Trace(W) is con¬
stant contribution (DC value) to the value of quadratic
form X^'^WX at all corners, X of unit hypercube. Thus,
the location of stable/ anti-stable states is invariant under
modification of Trace(W) value. Thus, from the standpoint of
location/computation of stable/anti-stable states, Trace(W) can
be set to zero.
■ It should be kept in mind that the dynamics of OHNN
with the symmetric synaptic weight matrix W is summa¬
rized by Theorem liUlLe. either there is convergence to
stable state or cycle of length 2 is reached. But in the
case RHNN with non-symmetric synaptic weight matrix,
there may be no convergence or the cycles are of length
strictly larger than 2. In the Section-3, we summarize our
experimental efforts.
B. Novel Complex Recurrent Hopfield Neural Network
In research literature on artificial neural networks, there
were attempts by many researchers to propose Complex
Valued Neural Networks (CVNNs) in which the synaptic
weights, inputs are necessarily complex numbers [2, MGZ].
In our research efforts on CVNNs, we proposed and studied
one possible Ordinary Complex Hopfield Neural Network
(OCHNN) first discussed in [3, RaP], [5-12]. The detailed
descriptions of such an artificial neural network requires the
following concepts.
1) Complex Hypercube: Consider a vector of dimension ’N’
whose components assume values in the following set
(1 -f jT, 1 —jl, —1 + jT, —1 — jlj.Thus there are points
as the corners of a set called the Complex Hypercube.
2) Complex Signum Function: Consider a complex number
’a H- jb’. The ’Complex Signum Function’ is defined as
follows:
Csign{a -f jb) = Sign{a) -f jSign{b)
In the same spirit of the real valued RHNN, we briefly summa¬
rize Complex Recurrent Hopfield Neural Network (CRHNN):
• The complex synaptic weight matrix of such an artificial
neural network (CRHNN) need NOT be a Hermitian
matrix. The activation function in such a complex valued
neural network is the complex signum function. The
convergence Theorem associated with Ordinary Complex
Hopfield Neural Network (OCHNN) is provided in [3,
RaP]. We are experimentally investigating the dynamics
of CRHNN.
HI. Dynamics of Recurrent Hopfield Network:
Experiments
In this section, certain network configurations are simulated
with various initial conditions (lying on the unit hypercube)
in order to provide insight on the dynamics of the Recurrent
Hopfield Networks.
A. Topology of the networks
The recurrent neural networks whose dynamics are pre¬
sented in this study have asymmetric synaptic weight matrices
with zeros on the diagonal (i.e. no self loops are allowed. As
discussed in Remark(4), this assumption can be made without
loss of generality). Edge weights (the synaptic weights) are
allowed to assume both positive and negative signs. The size
of the network (i.e. number of neurons) directly determines
the size of synaptic weight matrix. In the experiments, we are
going to investigate three cases: synaptic weight matrices with
(i) both positive and negative, (ii) only positive and (iii) only
negative entries.
B. Initialization of the network
A Recurrent Hopfield Network, in general, can be randomly
initialized in a state s where
s G'^.
T* is the state space and
IT-I = 2".
The size of the state space exponentially increases as the
number of neurons are increased. In order to reduce the
0
11
13
-15
6
-8
-12
-20
0
5
-5
4
-16
15
5
-17
0
7
13
13
-6
10
8
-9
0
1
-19
10
4
19
17
-3
0
5
-8
-11
-20
9
5
-7
0
16
-12
4
2
1
-17
13
0
Fig. 1. Toy synaptic weight matrix
computational overhead, the number of initial states utilized
are limited to a maximum of 1024.
• The decimals numbers in the range [0,1024) are converted
to binary and then all O’s are mapped to -I’s (From this
point on, states are going to be referred by using their
decimal values).
C. Modes of operation in experiments
During the experiments, both parallel and serial modes are
tested. However, serial mode results in two specihc cases.
The hrst is the case where network converges and the second
is when there is a cycle of 2 as iterations proceed. On the
other hand, parallel mode exhibits more interesting outcomes.
Hence, the focus during the experiments have been placed on
this update mode.
D. Results
1) Convergence and cycles: Networks with various
polarities of synaptic connections behave expectedly in
different manners. The experiments have led to certain
types of observations in terms of convergence and cycle
characteristics. In vast majority of trials, it has been observed
that given an initial state from T*, the network either converges
to a single state or converges into a certain pattern of cycle
in multiple iterations.
■ In networks with both positive and negative synaptic
connections, depending on the network size, convergence
can take place immediately if the initial state is located in
a small proximity to one of the basins of attraction in the
energy landscape. However simple generalization of the
network dynamics is not possible. The network can be led
to certain cycles which are of various lengths. One main
observation is that a collection of distinct initial states can
be driven into one or more different cycles. This result is
interesting since it may imply that a Recurrent Hopheld
Network can be used as a function to map signals into
certain common cycles. To illustrate the possible state
transitions, the network in Fig[T] is to be used.
• When parallel mode of update is applied on this network,
the results revealed that there exist 6 cycles in which
the RHNN loops given all possible initializations. In this
mode, length of the cycles varied between 2 and 4 while
in the serial mode of operation, the network reached
cycles of length of at most two. Furthermore, it often
TABLE I
The cycles for the toy example
Mode
Cycles
Parallel
35, 58-69, 42-5, 21-106, 22-119-92-94, 85-122
Serial
11, 35, 36, 41, 45, 57, 61, 63, 64, 66, 70, 86,
91, 92, 100-12, 27-115, 118-112, 118-31
converged to certain minima points in state space. To
note, the cycles of length 1 are used to refer to the
stable states. The observed cycles, following the naming
convention provided previously, are listed in TABLE U
• If the non-symmetric synaptic weight matrix is non¬
negative or non-positive, it is observed that the network
converges to a single state or to a cycle of length 2.
• If the non-symmetric synaptic weight matrix has positive
and negative elements, it is observed that the cycles are
of length strictly equal or larger than 2.
« It is observed that in the case of various RHNNs, there are
more than one directed cycles in the state space (reached
with a change of initial conditions).
■ In some special cases a single cycle is reached.
• The simulations show that the networks of mixed-polarity
connections are favoring certain states or combinations of
the whole possible state space. The parallel mode yielded
that inputs that are close in terms of Hamming distance
usually go through the same cycles or end up at certain
stable states.
> Larger networks with more than 20 neurons are observed
to fail to reach at detectable patterns or single state in the
parallel mode within reasonable number of iterations.
Next section will focus on the energy values associated with
various states. The energy values may be useful to make better
conclusions on the observations.
E. Energy of the Network during cycles and convergence
The Recurrent Hopheld Networks experimented in this
study exhibit two characteristics. First characteristic is that
the network reaches a constant energy value when it gets
trapped in a single state or a set of states (a cycle). The second
characteristic is that the energy of the network Huctuates
around a certain DC value. The latter has been observed to
occur in cycles only. The typical energy states of the toy
example during 128 trials are given in hgures Figl2]- Figjh]
It is useful to notice that the network’s energy states (during
convergence or cyclic period) which are caused by a collection
of distinct initial states are the same recapitulating what has
been stated previously.
. InFiglU a common energy prohle of network is shown.
Usually, in the toy example, the network conhgures itself
into one of the low energy regions.
• FigO and Fig|5] shows that the network obtains a steady
energy even during the cycles while Fig|4] and Fig|6]
indicates that the network’s energy oscillates as well as
the state it assumes on each iteration.
^ ^ ^ . * 1 j • * 1 . 1 oc Fig. 5. Energy of the network during the cycle 21-106.
rig. 2. Energy or the network during the cycle 35. a oj o j
Energy of the toy network during iterations
Fig. 3. Energy of the network during the cycle 58-69.
Fig. 6. Energy of the network during the cycle 22-119-92-94.
Fig. 4. Energy of the network during the cycle 42-5 and 85-122.
• The observations indicate that larger networks exhibit
similar behavior in the parallel mode but with the ex¬
ception that energy fluctuations are more often present
during cyclic period.
• In serial mode of operation, the energy of the network
remains same if it undergoes a single state convergence
whereas it fluctuates if in a cycle of 2.
IV. Applications
Non-linear dynamical systems exhibit interesting dynamic
behavior such as (i) convergence, (ii) oscillations (periodic
behavior) and (iii) chaos.
Researchers have identified certain applications (such as
cryptography) for interesting non-linear dynamical systems.
Specifically neural circuits based on non-linear dynamical
systems were identified with certain functional capabilities of
biological brains. The authors realized that Recurrent Hopfield
Network in this research paper could be capitalized for various
applications.
• The main observation is that the cycles of various lengths
in the state space could be utilized for applications such
as ’’MULTI-PATTERN” based associative memory. Based
on the length of cycle, a memory utilizing multiple
patterns (corners of the hypercube) can be synthesized.
• In the RHNNs investigated, sometimes CONSENSUS to
a single cycle is achieved. Thus, such networks could be
utilized for memorizing a collection of states correspond¬
ing to a cycle.
• In some RHNNs simulated, multiple cycles are reached
when the initial condition is varied. Thus, in this case
CONSENSUS to multiple memory groups (of states) is
achieved.
V. Conclusion
In this research paper, our efforts have focused on in¬
troducing a novel real/complex valued Recurrent Hopfield
Neural Network. In addition, we have shown a method to
synthesize a desired energy landscape for such networks. In
the experimental section, the dynamics of such real valued
Recurrent Hopfield Networks have been investigated in terms
of the state transitions and energy. Finally, certain potential
applications which exploit the shown properties are proposed.
References
[1] J.J. Hopfield, “Neural Networks and Physical Systems
with Emergent Collective Computational Abilities,”
Proceedings of National Academy Of Sciences, USA
Vol. 79, pp. 2554-2558, 1982.
[2] Mehmet Kerem Muezzinoglu, Cuneyt Guzelis and
Jacek M. Zurada, “A new design method for the
Complex-valued Multistate Hopfield Associative
Memory,” IEEE Transactions on Neural Networks, Vol.
14, No. 4, July 2003.
[3] G. Rama Murthy and D. Praveen, “Complex-valued
Neural Associative Memory on the Complex
hypercube,” IEEE Conference on Cybernetics and
Intelligent Systems, Singapore, 1-3 December, 2004.
[4] G. Rama Murthy, “Optimal Signal Design for Magnetic
and Optical Recording Channels,” Bellcore Technical
Memorandum, TM-NWT-018026, April 1st , 1991.
[5] G. Rama Murthy and B. Nischal, “Hopfield-Amari
Neural Network : Minimization of Quadratic forms,”
The 6th International Conference on Soft Computing
and Intelligent Systems, Kobe Convention Center
(Kobe Portopia Hotel), November 20-24, 2012, Kobe,
Japan.
[6] Akira Hirose, “Complex Valued Neural Networks:
Theories and Applications,” World scientific publishing
Co, November 2003.
[7] G. Rama Murthy and D. Praveen, “A Novel Associative
Memory on the Complex Hypercube Lattice,” 16th
European Symposium on Artificial Neural Networks,
April 2008.
[8] G. Rama Murthy, “Some Novel Real/ Complex Valued
Neural Network Models,” Advances in Soft Computing,
Springer Series on Computational Intelligence: Theory
and Applications, Proceedings of 9th Euzzy days,
September 18-20 2006, Dortmund, Germany.
[9] G. Jagadeesh, D. Praveen and G. Rama Murthy,
“Heteroassociative Memories on the Complex
Hypercube,” Proceedings of 20th IJCAI Workshop on
Complex Valued Neural Networks, January 6-12 2007.
[10] V. Sree Hari Rao and G. Rama Murthy, “Global
Dynamics of a class of Complex Valued Neural
Networks,” Special issue on CVNNS of International
Journal of Neural Systems, April 2008.
[11] G. Rama Murthy, “Infinite Population, Complex Valued
State Neural Network on the Complex Hypercube,”
Proceedings of International Conference on Cognitive
Science, 2004.
[12] G. Rama Murthy, “Multidimensional Neural Networks-
Unified Theory,” New Age International Publishers,
New Delhi, 2007.