Artificial Intelligence and Quantum Computing for Advanced Wireless Networks. Savo G. Glisic

Artificial Intelligence and Quantum Computing for Advanced Wireless Networks

alt="sigma-summation Underscript i Endscripts upper R Subscript i left-arrow j Superscript left-parenthesis l comma l plus 1 right-parenthesis Baseline equals upper R Subscript j Superscript left-parenthesis l plus 1 right-parenthesis Baseline dot left-parenthesis 1 minus StartFraction b Subscript j Baseline Over z Subscript j Baseline EndFraction right-parenthesis"/>

where the multiplier accounts for the relevance that is absorbed (or injected) by the bias term. If necessary, the residual bias relevance can be redistributed onto each neuron x_i. A drawback of the propagation rule of Eq. (4.18) is that for small values z_j , relevances R_{j ← j} can take unbounded values. Unboundedness can be overcome by introducing a predefined stabilizer ε ≥ 0:

(4.20)

The conservation law then becomes

(4.21)

where we can observe that some further relevance is absorbed by the stabilizer. In particular, relevance is fully absorbed if the stabilizer ε becomes very large.

An alternative stabilizing method that does not leak relevance consists of treating negative and positive pre‐activations separately. Let z Subscript j Superscript plus Baseline equals sigma-summation Underscript i Endscripts z Subscript italic i j Superscript plus Baseline plus b Subscript j Superscript plus and z Subscript j Superscript minus Baseline equals sigma-summation Underscript i Endscripts z Subscript italic i j Superscript minus Baseline plus b Subscript j Superscript minus Baseline comma where – and + denote the negative and positive parts of z_ij and b_j . Relevance propagation is now defined as

(4.22)

where α + β = 1. For example, for αβ = 1/2, the conservation law becomes

(4.23)

which has a similar form as Eq. (4.19). This alternative propagation method also allows one to manually control the importance of positive and negative evidence, by choosing different factors α and β.

Once a rule for relevance propagation has been selected, the overall relevance of each neuron in the lower layer is determined by summing up the relevances coming from all upper‐layer neurons in agreement with Eqs. (4.8) and (4.9):

(4.24) upper R Subscript i Superscript left-parenthesis l right-parenthesis Baseline equals sigma-summation Underscript j Endscripts upper R Subscript i left-arrow j Superscript left-parenthesis l comma l plus 1 right-parenthesis

Schematic illustration of relevance propagation (heat map; relevance is presented by the intensity of the red color).

Figure 4.3 Relevance propagation (heat map; relevance is presented by the intensity of the red color).

Source: Montavon et al. [92].

The relevance is backpropagated from one layer to another until it reaches the input pixels x_(d), and where the relevances upper R Subscript d Superscript left-parenthesis 1 right-parenthesis provide the desired pixel‐wise decomposition of the decision f(x). A practical example of relevance propagation obtained by Deep Taylor decomposition is shown in Figure 4.3 [92].

4.3 Rule Extraction from LSTM Networks

In this section, we consider long short term memory networks (LSTMs), which were discussed in Chapter 3, and described an approach for tracking the importance of a given input to the LSTM for a given output. By identifying consistently important patterns of words, we are able to distill state‐of‐the‐art LSTMs on sentiment analysis and question answering into a set of representative phrases. This representation is then quantitatively validated by using the extracted phrases to construct a simple rule‐based classifier that approximates the output of the LSTM.

Word importance scores in LSTMS: Here, we present a decomposition of the output of an LSTM into a product of factors, where each term in the product can be interpreted as the contribution of a particular word. Thus, we can assign importance scores to words according to their contribution to the LSTM’s prediction. We have introduced the basics of LSTM networks in the Chapter 3. Given a sequence of word embeddings x₁, x_T ∈ ℝ^d, an LSTM processes one word at a time, keeping track of cell and state vectors (c₁, h₁), (c_T, h_T), which contain information in the sentence up to word i. h_t and c_t are computed as a function of x_t, c_{t − 1} using the updates given by Eq.

Скачать книгу