Материал: искусственный интеллект

2 Related Work

A composite system being entangled cannot be validly decomposed and modeled as separate subsystems. The quantum theory provides formal tools to model interacting systems as non-decomposable in macroscopic world as well. The phenomenon of quantum entanglement has been investigated in semantic spaces making use of Hyperspace Analogue to Language (HAL) model [13, 14]. Hou et al. considered high order entanglements that cannot be reduced to the compositional e ect of lower-order ones, as an indicator of high-level semantic entities. Melucci proposes quantum-like entanglement for modeling the interaction between a user and a document as a composite system [15].

The non-compositionality of entangled systems opened also the door to developing quantum-like models of cognitive phenomena which are not decompositional in nature. Concept combinations have been widely modeled as composite systems [1, 2, 6, 7, 22]. The state of the composite system between two words can be modeled by taking the tensor product of the states of the individual words respectively. If the concept combination is factorizable, then the concept is compositional in the sense it can be expressed as a product of states corresponding to the separate words. A concept that is not factorizable cannot be expressed by either the first or the second word individually, and is deemed non-compositional, and termed entangled [7].

Quantum theory provides a well-developed set of analytical tools that can be used to determine whether the state of a system of interest can be validly decomposed into separate sub-systems. A possible way to test the non-compositional state of a composite system is the violation of Bell’s inequalities. For instance, having calculated the expectation values of variables associated with an experiment, we can fit the Clauser-Horne-Shimony-Holt (CHSH) version of Bell’s inequality [9]. If the CHSH inequality is greater than 2, then the Bell inequality

is violated. It has been empirically found that the maximal possible violation in

√

quantum theory is 2 2 ≈ 2.8284 [8]. This means that each violation being close to the maximal value is very significant. In addition to the CHSH inequality, Bruza et al. [7] propose Clauser-Horne inequalities to analyse the decomposability of quantum systems. The Schmidt decomposition is another way for detecting entanglement in bipartite systems [17]. According to the theorem, after decomposition, each pure state of the tensor product space can be expressed as the product of subsystem orthonormal bases and non-negative real coe cients. The square sum of the coe cients is equal to 1. The number of non-zero coe cients is called Schmidt number. If it equals 1, then the composite state is the product state. If it greater than 1, then the composite state is non-compositional.

So far, researchers have used joint probabilities in cognitive science for calculating expectation values assuming that the outcomes of observables are dependent. Additionally, probabilities can be calculated via trace formula in Gleason’s theory [11]. In a similar way, expectation value of two random variables is defined the product of traces [15]. Finally, probabilities could be re-expressed as function of an angle θ, where θ is defined as a di erence in phase between two random

168 D. Gkoumas et al.

observables, once we view the relationship between them as a geometrical relationship [15].

3 Quantum Entanglement and Bell Inequality

Let us suppose that we have a system of two qubits expressed in a Bit basis {0, 1}, such that the first qubit is in a state a0|0 + a1|1 and the second one in a state b0|0 + b1|1 . The state of the two qubits together as a composite system is a superposition of four classical probabilities resulting in

Let us now assume that the composite system is in an entangled state given by the following Bell state

22

When we measure the composite system, the probability of the system to collapse either to the state |00 or to the state |11 is equal to 0.5. However, after a measurement, the system is not in an entangled state anymore. For instance, once we measure the state |00 , the new state of the system results in

Let us now assume that we measure the state |0 of the first qubit (Eq. (2)). Then the probability for the first qubit to collapse to the state |0 again equals 0.5. However, after the measurement, the probability of the second qubit to be in the state |0 currently equals 1. Let us suppose that we change the Bit basis to a Sign basis {−, +}. According to the rotation invariance [19], the Bell state in the Sign basis is again an equal superposition of the state | − − and the state | + + such that

22

Suppose now that we want to measure the probability of the second qubit to be in the state |− according to the Sign basis, given that we have already measured the probability of the first qubit to be in state |0 concerning the Bit basis. Once we measure the first qubit, the probability of the second qubit to be in the same state |0 is equal to 1. If θ is the angle between the Bit and Sign basis, then according to the Pythagorean theorem, the probability of the second qubit to be in the state |− equals cos2 θ.

In quantum mechanics, the criteria used to test entanglement are given by Bell’s inequalities. A possible way to proceed is to define four observables. Each observable has binary values ±1 thus give two mutually exclusive outcomes. For

instance, a photon can be detected by ‘+’ or ‘−’ channel (see Fig. 9). Let us denote as A1, B1 the observables describing the first system, and A2, B2 the observables of the second one. If a composite system is separable, the following CHSH inequality holds:

where denotes the expectation value between two observables. The calculation of expectation values will be articulated in Sect. 4. The violation of (5) is a sign of entanglement. A Bell inequality violation implies that at least one of the assumptions of local-realism made in the proof of (5) must be incorrect [16]. This points to the conclusion that either or both of locality - an object is only directly influenced by its immediate surroundings - and realism - an object has definite values - must be rejected as a property of the composite systems violating CHSH inequality.

4Non-classical Correlations in Decision Fused Multimodal Documents

Before the late fusion process, there exists a probability p(R|T ) for a multimodal document DM to be relevant to a multimodal information need concerning the textual information. Similarly, the probability for the same document not to be relevant is denoted as p(R|T ), which is equal to 1−p(R|T ). Let us consider a realvalued two dimensional Hilbert Space for the relevance of the DM concerning the textual information (Fig. 6). In Fig. 6 the vector Rt stands for the relevance of the document concerning the text-based modality. On the other hand, the Rt represents the non-relevance with respect to the same text-based information need and is orthogonal to Rt.

The text-based relevance of a document can be modeled as a vector in the Hilbert Space, which unifies the logical, probabilistic and vector space based approaches to IR [21]. This vector is a superposition of relevance and nonrelevance vectors with respect to the text-based modality and is represented as:

where |a|2 + |a |2 = 1. The coe cients a and a are captured by taking the projection of |DM onto the relevance and non-relevance vectors respectively

In a similar way, we denote as p(R|I) the probability for a multimodal document DM to be relevant concerning the image-based information need, and

p(R|I) the probability to be irrelevant respectively (Fig. 7). The relevance of a document with respect to the image-based modality can be in a similar manner modeled as:

170 D. Gkoumas et al.

In this case, p(R|I) is computed as the square of the inner product | Ri|DM |2. Likewise, p(R|I) equals to | Ri|DM |2.

After the late fusion process, the document can be judged based on both textbased and image-based modalities. Such a phenomenon can be modeled in the same Hilbert Space by having a di erent basis for each modality, as presented in Fig. 8. The document DM is represented as a unit vector and its representation

is expressed with respect to the bases T = {|Rt , |Rt } and I = {|Ri , |Ri } fusing at the end the local decisions. Each basis models context with respect to a given modality.

The rest of the experimental setup is analogous to that one for investigating quantum entanglement in photons [3]. Figure 9 shows the experimental setup for the violation of Bell’s inequalities. The source S produces a pair of photons, sent in opposite directions. Each photon encounters a two-channel polariser whose orientation can be set by the experimenter. Coincidences (simultaneous detections) are recorded, the results being categorised as ++, +−, −+, or − − and corresponding counts accumulated by the coincidence monitor CM.

Fig. 9. Schematic of a “two-channel” Bell test

Now let us consider Fig. 10, which depicts two multimodal documents, DM 1 and DM 2 respectively. As afore-mentioned, the documents DM 1 and DM 2 can be expressed with either the text-based basis or image-based basis being in

Fig. 10. The interaction between two documents is modeled as a composite system

a superposition of relevance and non-relevance states. In quantum theory, the interaction between DM 1 and DM 2 can be modeled as a composited system by using the tensor product of the document Hilbert Spaces. The state of the composite system DM 1 DM 2 can be obtained by taking the tensor product of the relevance and non-relevance states. Concerning the text-based modality, the state of the composite system is defined as follows:

|DM 1 |DM 2 = (a1|Rt + a1|Rt ) (a2|Rt + a2|Rt )

(8)

= a1a2|RtRt + a1a2|RtRt + a1a2|RtRt + a1a2|RtRt ,

where |a1a2|2 + |a1a2|2 + |a1a2|2 + |a1a2|2 = 1. In a similar way, if we define the image-based basis as a standard basis then the state of the composite system

DM 1 DM 2 concerning the image-based modality can be expressed as follows:

|DM 1 |DM 2 = (b1|Ri + b1|Ri ) (b2|Ri + b2|Ri )

(9)

= b1b2|RiRi + b1b 2|RiRi + b1b2|RiRi + b1b2|RiRi

In Eq. (8), the first and second terms reveal that when the text-based content of the DM 1 is relevant, then we cannot be sure about the relevance of the textbased content of the other document. Similarly, according to the third and fourth term in Eq. (8), when the text-based content of the DM 1 is non-relevant, then the other document is in a superposition of relevance and non-relevance states with respect to the text-based modality. The same is observed when we consider the image-based basis as a standard basis.

If the state of the text-based (Eq. (8)) or image-based (Eq. (9)) composite system is factorizable, then the system is compositional in the sense it can be expressed as a product of states corresponding to the separate subsystems. A composite quantum system that is not factorizable is deemed non-compositional and termed entangled [7]. In the last case, if we consider the text-based representation as a standard basis, then we can define two Bell states, either the state