Beruflich Dokumente
Kultur Dokumente
Adenilton Silva
, Wilson de Oliveira
.
.
.
.
.
. y
.
.
. 00 . . . 1 C[1]
x
n
00 . . . 0 C[0]
Figure 1. Ram Node
A Probabilistic Logic Node (PLN) differs from a RAM
node in that a 2-bit number (rather than a single bit) is now
stored at the addressed memory location. The content of
this location is turned into the probability of ring (i.e.,
generating 1) at the overall output of the node. In other
words, a PLN consists of a RAM node, where now a
2-bit number is stored at the addressed memory location
(0, 1 or u), respectively represented as say 00, 11 and 01
(or 10), with a probabilistic output generator. The output of
the PLN Node is given by [10]:
y =
_
_
0, if C[x] = 0
1, if C[x] = 1
random(0, 1), if C[x] = u
The Multi-Valued Probabilistic Logic Node (MPLN) dif-
fers from PLN by allowing a wider but still discrete range
of probabilities to be stored at each memory content.
III. QUANTUM COMPUTATION
Quantum computing [11] was originally proposed by
Richard Feynman [12] in the 1980s and had its formalisation
with David Deutsch which proposed the quantum Turing
machine [13]. Quantum Computing has been popularised
through the quantum circuit model [14] which is a math-
ematical quantisation [15] of the classical boolean circuit
model of computation. Quantum computers, if ever built,
can also be seen as a parallel device for potentially improve
the computational efciency of neural networks [16].
The quantum information unit is the quantum bit or
qubit. A very intuitive view of the quantisation procedure
used almost everywhere is put forward by Nik Weaver in the
Preface of his book Mathematical Quantization [15] with
says in a nutshell: The fundamental idea of mathematical
quantisation is sets are replaced with Hilbert spaces. In or-
der to properly deal with operations we would add: functions
(morphisms) are replaced with linear operators in general -
and in some cases, with unitary operators in particular.
The quantisation of the boolean circuit logic starts by
simply embedding the classical bits {0, 1} in a convenient
Hilbert space. The natural way of doing this is to rep-
resent them as (orthonormal) basis of a Complex Hilbert
space. In this context these basis elements are called the
computational-basis states [17]. Linear combinations (from
Linear Algebra [18]) of the basis spans the whole space
whose elements, called states, are said to be in super-
position. The general state of a qubit is a superposition
(linear combinations) of the two computational-basis states:
| = |0 + |1, where , are complex coefcients
(called probability amplitudes) constrained by the normal-
isation condition: ||
2
+ ||
2
= 1; and |0, |1 are a pair
of orthonormal basis vectors representing each classical bit,
or cbit, as column vector. For example the cbits can be
represented as
|0 =
1
0
, |1 =
0
1
, | =
_
=
_
1
_
_
Which obviously generalises to any pair of n and
mdimensional vectors producing a nm-dimensional vec-
tor or more generally a (n
0
, m
0
)dimensional ma-
trix by a (n
1
, m
1
)dimensional one producing a third
(n
0
m
0
, m
0
n
0
)dimensional matrix and so on. Operations
on qubits are carried out by unitary operators. So quantum
algorithms on n bits are represented by unitary operators
U over the 2
n
-dimensional complex Hilbert space: |
U |.
In quantum computation, all operators on qubits are
reversible, except for the process called measurement, that
loses the information about the superposition of states. To
measure a general state | collapses (projects) it into either
the |0 state or the |1 state, with probabilities ||
2
or ||
2
respectively.
260
Others widely used operators and their corresponding
matrix operators are [11]:
I, identity operator: which is particularly useful when
combined with other in the case one wants to manip-
ulate a particular bit in a register leaving the others
intact.
I =
1 0
0 1
I |0 = |0
I |1 = |1
X, ip operator: behaves as the classical NOT on the
computational basis.
X =
0 1
1 0
X|0 = |1
X|1 = |0
H, Hadamard transformation: generates superposition
of states.
H =
1
1 1
1 1
H|0 = 1/
2(|0 + |1)
H|1 = 1/
2(|0 |1)
W, operator used in the PQM node.
W =
e
i
2n
0
0 1
H|0 = 1/
2(|0 + |1)
H|1 = 1/
2(|0 |1)
Quantum operators also may be represented in quantum
circuits by corresponding quantum gates. Figure 2
shows an n-qubit controlled gate U whose action on
the target qubit (bottommost) is active or not by n 1
(topmost) control qubits [11]. The output is checked by
measurement gates.
.
.
..
.
.
where
=
X X
Figure 2. A quantum circuit
IV. QUANTUM NEURAL NETWORKS
The concept of quantum neural computation was rst intro-
duced by Kak in 1995, creating a new paradigm using neural
networks and quantum computation which opens several
new directions in neural network research [19]. It is ex-
pected that quantum neural networks are more efcient than
classical neural networks, parallel to what is expected from
quantum computation in relation to Classical Computation.
In that same year, Menneer and Narayanan proposed
a Quantum Inspired Neural Network applied the multiple
universes view from quantum theory to one-layer articial
neural networks [20]. In 1998 Ventura and Martinez in-
troduced a quantum associative memory with a capacity
exponential in the number of neurons [21].
The non-linear activation functions used in models of neu-
ral networks delayed further development of their quantum
analogue. But in 2001 Altaisky proposed a quantum system
where the rules of perceptron learning were proposed [22].
Several Quantum Weighted Neural Networks models have
been proposed but there remains the challenge of direct
implementation in quantum circuits, natural adaptation of
the learning algorithms and quantum learning algorithms
respecting the postulates of Quantum Mechanics. These are
characteristics not altogether found in any of the proposed
Quantum Weighted Neural Networks models. The rst really
quantum neural networks is the weightless neural networks
based on quantum RAM in [1].
V. PROBABILISTIC QUANTUM MEMORIES
The probabilistic quantum memory PQM was proposed by
Trugenberger in 2001 [7]. The advantage of the PQM mem-
ory is its exponential capacity of storage. In [8] Trugenberger
use this memory to dene a quantum associative memory
taking beneces of this memory capacity.
In the PQM memory the information retrieval step seems
to contradict the general idea of a memory [23], since after
the retrieval step the memory needs to be prepared again.
Trungerberg argues that its model is efcient if the number
of inputs is polynomial [24]. The trouble with the PQM
memory is the need to measure the memory register in order
to recover the data stored in the memory.
The PQM memory is a state which represents a super-
position of the p patterns p
j
to be stored, as showed in
equation (1). The algorithm to create this state is described
in [7] with details; With this algorithm 2
n
patterns with n
qubits can be stored in n qubits.
|m =
1
p
p
j=1
p
j
(1)
The retrieval algorithm of the PQM requires three regis-
ters: the rst i will receive the input pattern, the second m
will contain the memory |m and the third c is a auxilary
quantum bit |c initializated as H|0.
|
0
=
1
2p
p
k=1
i
1
, , i
n
; p
k
1
; ; p
k
n
; 0
+
1
2p
p
k=1
i
1
, , i
n
; p
k
1
; ; p
k
n
; 1
(2)
After the deterministic part of the retrieval algorithm, the
quantum state will be as indicated in the equation (3), where
d
H
i, p
k
p
p
k=1
cos
2n
d
H
i, p
k
i
1
, , i
n
; p
k
1
; ; p
k
n
; 0
+
1
p
p
k=1
sin
2n
d
H
i, p
k
i
1
, , i
n
; p
k
1
; ; p
k
n
; 1
(3)
VI. PQM NODE
The measurement of the memory register in a PQM memory
is a negative characteristic since the memory collapse makes
it loose its future uses and one then need to prepare its state
again. In this section we propose a weightless neural node
based on the PQM memory and without the need to measure
the memory register.
In the RAM node the decision about the number of inputs
n
1
has directly inuence in the generalisation ability of the
network, however one of the main disadvantage of the RAM
node is that its memory grows exponentially with n
1
. In
this section will be dened a quantum node similar to the
RAM node, but with architecture based in the PQM memory
instead of the RAM memory, in such a way that the PQM
node will have the size of the memory equal to its number
of inputs.
A PQM node is implemented as a PQM memory. As in
the retrieval algorithm of the PQM memory, a PQM node
with n inputs has three registers: the rst is the register i with
n qubits in which we feed the input of the node; the second
memory register is m with n qubits to hold the memory of
the node |m and the third register is c with a single qubit,
the measure of c will determine the output of the node in a
probabilistic way. A PQM node with n inputs is represented
by a quantum state with three registers containing 2n + 1
qubits.
As the RAM node the PQM node is used to recognize only
one class, a new input pattern will make the node answer
positively if the pattern is in the class that the node was
trained for or otherwise negatively rejecting the pattern.
The learning algorithm of a PQM node is to create a
superposition of all the patterns in the training set |m in the
memory register m of the node. This process can be realized
using the storage algorithm of the memory PQM [7]. The
cost of the training is linear in the number n
2
of patterns in
the training set, the computational cost is (n
2
).
After feeding one new pattern in the input register of
the node, the deterministic part of the retrieval algorithm
of the PQM memory is applied to the state (Fig. 3), at
this moment the equation (3) describes the state of the
node. A measurement of the value in the memory register
c will determine the output of the node in a probabilistic
way described in the equation (4), where d
H
(i, p
k
) is the
Hamming distance between the input pattern and the kth
pattern stored in the memory.
P (|c = |0) =
p
k=1
1
p
cos
2
2n
d
H
(i, p
k
)
P (|c = |1) =
p
k=1
1
p
sin
2
2n
d
H
(i, p
k
)
(4)
For instance, if one class is represented in the training set
by {(00000000), (00000001)}, then it is used the algorithm
proposed in [7] to create the state described in the equation
(5).
| =
1
2
(|00000000 + |00000001) (5)
Feeding the pattern (11111111) in the input of the net-
work and initializing the register c with the state |c =
1
2
(|0 + |1) the state of the node at this moment is
described in the equation (6).
| =
1
2
[|11111111 (|00000000 + |00000001) |0 +
|11111111 (|00000000 + |00000001) |1]
(6)
After the action of the node, described in the Fig. 3 the
state | will be in the state described by the equation
| =
1
2
cos
8pi
16
|11111111 |00000000 |0 +
cos
7pi
16
|11111111 |00000001 |0 +
sin
8pi
16
|11111111 |00000000 |1 +
sin
7pi
16
|11111111 |00000001 |1
(7)
A measure in the register c of the node will result in |0,
the node recognize the pattern, with probability
2
cos
8pi
16
2
+
2
cos
7pi
16
2
= 1.9 10
2
(8)
and will result in |1 with probalility
2
sin
8pi
16
2
+
2
sin
7pi
16
2
= 98.1 10
2
(9)
and the pattern (11111111) will not be recognized.
The probabilities in the equations (8) and (9) can be
estimated repeating the execution of the node T times or
using a register c with more qbits [8].
Other disadvantage of the RAM node is the inability to
generalise; the ability of generalisation is achieved only with
RAM networks. As can be seen in the equation (4) the
output of a PQM node considers the Hamming distance,
in this equation the output |0 means that the pattern was
262
|i
1
|i
2
.
.
.
|i
n
|m
1
X W W
2
X
|m
2
X W W
2
X
.
.
.
W W
2
|m
n
X W W
2
X
|c H
Figure 3. PQM node
recognized, the output |0 will occur with high probability
if the input is near to the values stored in |m and the output
equal to |1 means that the pattern was not recognized, this
will occur if the input is distant to the values in |m. A new
pattern |i
, p
k
)
are near to zero. On the other hand, the storage of a pattern
|p