Skip to main content
Log in

Solution Structure of Multi-layer Neural Networks with Initial Condition

  • Published:
Journal of Dynamics and Differential Equations Aims and scope Submit manuscript

Abstract

This paper studies the initial value problem of multi-layer cellular neural networks. We demonstrate that the mosaic solutions of such system is topologically conjugated to a new class in symbolic dynamical systems called the path set (Abram and Lagarias in Adv Appl Math 56:109–134, 2014). The topological entropies of the solution, output, and hidden spaces of a multi-layer cellular neural network with initial condition are formulated explicitly. Also, a sufficient condition for whether the mosaic solution space of a multi-layer cellular neural network is independent of initial conditions is addressed. Furthermore, two spaces exhibit identical topological entropy if and only if they are finitely equivalent.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Fig. 1
Fig. 2
Fig. 3
Fig. 4
Fig. 5
Fig. 6
Fig. 7
Fig. 8
Fig. 9
Fig. 10
Fig. 11

Similar content being viewed by others

References

  1. Abram, W., Lagarias, J.C.: p-Adic path set fractals and arithemetics. J. Fractal Geom. 1, 45–81 (2014)

    Article  MathSciNet  MATH  Google Scholar 

  2. Abram, W., Lagarias, J.C.: Path sets in one-sided symbolic dynamics. Adv. Appl. Math. 56, 109–134 (2014)

    Article  MathSciNet  MATH  Google Scholar 

  3. Adler, R.L., Marcus, B.: Topological entropy and equivalence of dynamical systems. Mem. Am. Math. Soc., vol. 219. AMS: Providence, RI (1979)

  4. Arena, P., Baglio, S., Fortuna, L., Manganaro, G.: Self-organization in a two-layer CNN. IEEE Trans. Circuits Syst. I Fundam. Theory Appl. 45, 157–162 (1998)

    Article  Google Scholar 

  5. Arik, S., Tavsanoglu, V.: Equilibrium analysis of non-symmetric cnns. Int. J. Circ. Theor. Appl. 24, 269–274 (1996)

    Article  Google Scholar 

  6. Ban, J.-C., Chang, C.-H.: On the monotonicity of entropy for multi-layer cellular neural networks. Int. J. Bifur. Chaos Appl. Sci. Eng. 19, 3657–3670 (2009)

    Article  MathSciNet  MATH  Google Scholar 

  7. Ban, J.-C., Chang, C.-H.: On the structure of multi-layer cellular neural networks. Part II: The complexity between two layers (2012, under review)

  8. Ban, J.-C., Chang, C.-H.: Diamond in multi-layer cellular neural networks. Appl. Math. Comput. 222, 1–12 (2013)

    Article  MathSciNet  MATH  Google Scholar 

  9. Ban, J.-C., Chang, C.-H.: The learning problem of multi-layer neural networks. Neural Netw. 46, 116–123 (2013)

    Article  MATH  Google Scholar 

  10. Ban, J.-C., Chang, C.-H., Lin, S.-S.: The structure of multi-layer cellular neural networks. J. Differ. Equ. 252, 4563–4597 (2012)

    Article  MathSciNet  MATH  Google Scholar 

  11. Ban, J.-C., Chang, C.-H., Lin, S.-S., Lin, Y.-H.: Spatial complexity in multi-layer cellular neural networks. J. Differ. Equ. 246, 552–580 (2009)

    Article  MathSciNet  MATH  Google Scholar 

  12. Carmona, R., Jimenez-Garrido, F., Dominguez-Castro, R., Espejo, S., Rodriguez-Vazquez, A.: CMOS realization of a 2-layer cnn universal machine chip. In: Proceedings of the 2002 7th IEEE International Workshop on Cellular Neural Networks and Their Applications, 2002, (CNNA 2002), pp. 444–451 (2002)

  13. Chua, L.O.: Cnn: A Paradigm for Complexity, World Scientific Series on Nonlinear Science, Series A, vol. 31. World Scietific, Singapore (1998)

    Google Scholar 

  14. Chua, L.O., Yang, L.: Cellular neural networks: applications. IEEE Trans. Circuits Syst. 35, 1273–1290 (1988)

    Article  MathSciNet  Google Scholar 

  15. Chua, L.O., Yang, L.: Cellular neural networks: theory. IEEE Trans. Circuits Syst. 35, 1257–1272 (1988)

    Article  MathSciNet  MATH  Google Scholar 

  16. Chua, L.O., Shi, B.E.: Multiple layer cellular neural networks: a tutorial. In: Deprettere, E.F., van der Veen, A.-J. (eds.) Algorithms and Parallel VLSI Architectures, pp. 137–168. Elsevier, Amsterdam (1991)

    Google Scholar 

  17. Chua, L.O., Roska, T.: Cellular Neural Networks and Visual Computing. Cambridge University Press, Cambridge, MA (2002)

    Book  Google Scholar 

  18. Crounse, K.R., Chua, L.O.: Methods for image processing and pattern formation in cellular neural networks: a tutorial. IEEE Trans. Circuits Syst. 42, 583–601 (1995)

    Article  MathSciNet  Google Scholar 

  19. Crounse, K.R., Roska, T., Chua, L.O.: Image halftoning with cellular neural networks. IEEE Trans. Circuits Syst. 40, 267–283 (1993)

    Article  MATH  Google Scholar 

  20. Forti, M., Tesi, A.: New conditions for global stability of neural networks with application to linear and quadratic programming problems. IEEE Trans. Circuits Syst. I. Funda. Theory Appl. 42, 354–366 (1995)

    Article  MathSciNet  MATH  Google Scholar 

  21. Forti, M., Tesi, A.: A new method to analyze complete stability of PWL cellular neural networks. Int. J. Bifur. Chaos Appl. Sci. Eng. 11, 655–676 (2001)

    Article  MathSciNet  MATH  Google Scholar 

  22. Fukushima, K.: Artificial vision by multi-layered neural networks: neocognitron and its advances. Neural Netw. 37, 103–119 (2013)

    Article  MathSciNet  Google Scholar 

  23. Fukushima, K.: Training multi-layered neural network neocognitron. Neural Netw. 40, 18–31 (2013)

    Article  MATH  Google Scholar 

  24. Juang, J., Lin, S.-S.: Cellular neural networks: mosaic pattern and spatial chaos. SIAM J. Appl. Math. 60, 891–915 (2000)

    Article  MathSciNet  MATH  Google Scholar 

  25. Li, X.: Analysis of complete stability for discrete-time cellular neural networks with piecewise linear output functions. Neural Comput. 21, 1434–1458 (2009)

    Article  MathSciNet  MATH  Google Scholar 

  26. Liang, X.-B., Si, J.: Global exponential stability of neural networks with globally lipschitz continuous activations and its application to linear variational inequality problem. IEEE Trans. Neural Netw. 12, 349–359 (2001)

    Article  Google Scholar 

  27. Lind, D., Marcus, B.: An Introduction to Symbolic Dynamics and Coding. Cambridge University Press, Cambridge, MA (1995)

    Book  MATH  Google Scholar 

  28. Murugesh, V.: Image processing applications via time-multiplexing cellular neural network simulator with numerical integration algorithms. Int. J. Comput. Math. 87, 840–848 (2010)

    Article  MathSciNet  MATH  Google Scholar 

  29. Parry, W.: A finitary classification of topological markov chains and sofic systems. Bull. Lond. Math. Soc. 9, 86–92 (1977)

    Article  MathSciNet  MATH  Google Scholar 

  30. Peng, J., Zhang, D., Liao, X.: A digital image encryption algorithm based on hyper-chaotic cellular neural network. Fund. Inform. 90, 269–282 (2009)

    MathSciNet  Google Scholar 

  31. Savaci, F.A., Vandewalle, Joos, On the stability analysis of cellular neural networks. In: Proceedings, Second International Workshop on Cellular Neural Networks and their Applications, 1992, CNNA-92. IEEE, 1992, pp. 240–245 (1992)

  32. Takahashi, N., Chua, L.O.: On the complete stability of nonsymmetric cellular neural networks. IEEE Trans. Circuits-I. 45, 754–758 (1998)

    Article  MathSciNet  MATH  Google Scholar 

  33. Török, L., Roska, T.: Stability of multi-layer cellular neural/nonlinear networks. Int. J. Bifur. Chaos Appl. Sci. Eng. 14, 3567–3586 (2004)

    Article  MATH  MathSciNet  Google Scholar 

  34. Xavier-de Souza, S., Yalcin, M.E., Suykens, J.A.K., Vandewalle, J.: Toward CNN chip-specific robustness. IEEE Trans. Circuits Syst. I Reg. Papers 51, 892–902 (2004)

    Article  Google Scholar 

  35. Yang, Z., Nishio, Y., Ushida, A.: A two layer cnn in image processing applications. In: Proceedings of the 2001 International Symposium on Nonlinear Theory and Its Applicationsm, pp. 67–70 (2001)

  36. Yang, Z., Nishio, Y., Ushida, A.: Image processing of two-layer CNNs \(-\) applications and their stability. IEICE Trans. Fundam. E85–A, 2052–2060 (2002)

    Google Scholar 

Download references

Acknowledgments

The authors wish to express their gratitude to the anonymous referees for their careful reading and useful suggestions, which make significant improvements to this work. Ban is partially supported by the National Science Council, ROC (Contract No. NSC 102-2628-M-259-001-MY3). Chang is grateful for the partial support of the National Science Council, ROC (Contract No. NSC 104-2115-M-390-004-).

Author information

Authors and Affiliations

Authors

Corresponding author

Correspondence to Chih-Hung Chang.

Appendix: Equivalent Relation for Templates

Appendix: Equivalent Relation for Templates

The essential study of (1) is investigating two-layer cellular neural networks with the nearest neighborhood; namely, \(n = 2\) and \(\mathcal {N} = \{-1, 0, 1\}\). For the clarification of the investigation, we assume that \(a_{-1}^{(1)} = a_{-1}^{(2)} = 0\). Under such condition, (1) is represented as

$$\begin{aligned} \left\{ \begin{aligned} \dfrac{d}{dt} x_i^{(2)}(t)&= - x_i^{(2)}(t) + z^{(2)} + a^{(2)} y_i^{(2)}(t) + a_r^{(2)} y_{i+1}^{(2)}(t) + b^{(2)} y_i^{(1)}(t) + b_r^{(2)} y_{i+1}^{(1)}(t), \\ \dfrac{d}{dt} x_i^{(1)}(t)&= - x_i^{(1)}(t) + z^{(1)} + a^{(1)} y_i^{(1)}(t) + a_r^{(1)} y_{i+1}^{(1)}(t), \end{aligned} \right. \end{aligned}$$
(23)

where \(i \in \mathbb {N}\), \(t \ge 0\), and \(y^{(j)}_i = f(x^{(j)}_i)\) for \(j = 1, 2\). Since, for all ij, \(|x^{(j)}_i(t)| > 1\) provided t is large enough, the output \(y^{(j)}_i(t)\) is either 1 or \(-1\) after finite time. Hence we omit the time factor in the following discussion.

Suppose \(\mathbf {y} = \left( \begin{array}{ll} y_1^{(2)} y_2^{(2)} y_3^{(2)} \cdots \\ y_1^{(1)} y_2^{(1)} y_3^{(1)} \cdots \end{array} \right) \in \mathbf {Y}\) is a mosaic pattern. For \(i \in \mathbb {N}\), \(y_i^{(1)} = 1\) if and only if \(x_i^{(1)} > 1\). This derives

$$\begin{aligned} a^{(1)} + z^{(1)} - 1 > -a^{(1)}_r y^{(1)}_{i+1}. \end{aligned}$$
(24)

Similarly, \(y_i^{(1)} = -1\) if and only if \(x_i^{(1)}< -1\). This implies \(y_i^{(1)} = -1\) if and only if

$$\begin{aligned} a^{(1)} - z^{(1)} - 1 > a^{(1)}_r y^{(1)}_{i+1}. \end{aligned}$$
(25)

The same argument asserts

$$\begin{aligned} a^{(2)} + z^{(2)} - 1&> -a^{(2)}_r y^{(2)}_{i+1} - (b^{(2)} y^{(1)}_i + b^{(2)}_r y^{(1)}_{i+1}) \end{aligned}$$
(26)

and

$$\begin{aligned} a^{(2)} - z^{(2)} - 1&> a^{(2)}_r y^{(2)}_{i+1} + (b^{(2)} y^{(1)}_i + b^{(2)}_r y^{(1)}_{i+1}) \end{aligned}$$
(27)

are the necessary and sufficient condition for \(y_i^{(2)} = -1\) and \(y_i^{(2)} = 1\), respectively. Define \(\xi _1: \{-1, 1\} \rightarrow \mathbb {R}\) and \(\xi _2: \{-1, 1\}^{\mathbb {Z}_{3 \times 1}} \rightarrow \mathbb {R}\) by

$$\begin{aligned} \xi _1(w) = a^{(1)}_r w, \qquad \xi _2(w_1, w_2, w_3) = a^{(2)}_r w_1 + b^{(2)} w_2 + b^{(2)}_r w_3. \end{aligned}$$

Set

That is,

Since two-layer cellular neural networks are locally coupled systems, \(\mathcal {B}^{(1)}\) and \(\mathcal {B}^{(2)}\) represents the basic sets of admissible local patterns of the first and second layer of (23), respectively. The set of admissible local patterns \(\mathcal {B}\) of (23) is then

This is to say, the investigation of the equivalent relation for the templates is identical to the discussion of the equivalent relation for the basic sets of admissible local patterns. The fact that \(y^{(1)}, y_r^{(1)} \in \{-1, 1\}\) indicates \(a^{(1)} + z^{(1)} - 1 = - \xi _1(y_r^{(1)})\) and \(a^{(1)} + z^{(1)} - 1 = \xi _1(y_r^{(1)})\) partition \(a^{(1)}\)-\(z^{(1)}\) plane into 9 regions. More precisely, \(a^{(1)}\)-\(z^{(1)}\) plane is partitioned by

$$\begin{aligned} a^{(1)} + z^{(1)} - 1 = a^{(1)}_r, \quad a^{(1)} + z^{(1)} - 1 > - a^{(1)}_r, \end{aligned}$$

and

$$\begin{aligned} a^{(1)} - z^{(1)} - 1 = a^{(1)}_r, \quad a^{(1)} - z^{(1)} - 1 = - a^{(1)}_r. \end{aligned}$$

Encode these nine regions as [pq] for \(0 \le p, q \le 2\), then a pair of parameters \((a^{(1)}, z^{(1)}) \in [p, q]\) infers that \((a^{(1)}, z^{(1)})\) satisfies m inequalities in (24) and n inequalities in (25). The relative positions of

$$\begin{aligned} a^{(1)} + z^{(1)} - 1 = a^{(1)}_r \quad \text {and} \quad a^{(1)} + z^{(1)} - 1 > - a^{(1)}_r \end{aligned}$$

and the relative positions of

$$\begin{aligned} a^{(1)} - z^{(1)} - 1 = a^{(1)}_r \quad \text {and} \quad a^{(1)} - z^{(1)} - 1 = - a^{(1)}_r \end{aligned}$$

remain to be determined. The “order” of lines \(a^{(1)} + z^{(1)} - 1 = (-1)^{\ell } \xi _1(y_r^{(1)})\), \(\ell = 0, 1\), in the plane come from the sign of \(a_r^{(1)}\), this demonstrates that the parameter space \(\{(a^{(1)}, a^{(1)}_r, z^{(1)})\}\) is partitioned into \(2 \times 9 = 18\) equivalent regions. (Notably, the order of lines \(a^{(1)} - z^{(1)} - 1 = \xi _1(y_r^{(1)})\) and \(a^{(1)} - z^{(1)} - 1 = - \xi _1(y_r^{(1)})\) is determined seamlessly once the order of \(a^{(1)} + z^{(1)} - 1 = (-1)^{\ell } \xi _1(y_r^{(1)})\) is given.) Namely, any two sets of parameters located in the same region determine the identical basic set of admissible local patterns \({{\mathcal {B}}}^{(1)}\). See Fig. 12a.

In an analogous manner, \(y^{(2)}, y_r^{(2)}, u^{(2)}, u_r^{(2)} \in \{-1, 1\}\) indicates that \(a^{(2)} + z^{(2)} - 1 > - \xi _2(y_r^{(2)}, u^{(2)}, u_r^{(2)})\) and \(a^{(2)} + z^{(2)} - 1 > \xi _2(y_r^{(2)}, u^{(2)}, u_r^{(2)})\) partition \(a^{(2)}\)-\(z^{(2)}\) plane into 81 regions. Encode these regions as [pq] for \(0 \le p, q \le 8\), then a pair of parameters \((a^{(2)}, z^{(2)}) \in [p, q]\) infers that \((a^{(2)}, z^{(2)})\) satisfies m inequalities in (26) and n inequalities in (27). Furthermore, the order of \(a^{(2)} + z^{(2)} - 1 = \xi _2(y_r^{(2)}, u^{(2)}, u_r^{(2)})\) can be uniquely determined according to the following procedures.

  1. (1)

    The signs of \(a_r^{(2)}, b^{(2)}, b_r^{(2)}\).

  2. (2)

    The magnitude of \(a_r^{(2)}, b^{(2)}, b_r^{(2)}\).

  3. (3)

    The competition between the parameter with the largest magnitude and the others. In other words, suppose \(m_1 > m_2 > m_3\) represent \(|a_r^{(2)}|, |b^{(2)}|, |b_r^{(2)}|\). We need to determine whether \(m_1 > m_2 + m_3\) or \(m_1 < m_2 + m_3\).

This partitions the parameter space \(\{(a^{(2)}, a^{(2)}_r, b^{(2)}, b^{(2)}_r, z^{(2)})\}\) into \(8 \times 6 \times 2 \times 81 = 7776\) regions and each region is associated with a basic set of admissible local patterns (cf. Fig. 12b).

Fig. 12
figure 12

The partition of \(a^{(1)}\)-\(z^{(1)}\) and \(a^{(2)}\)-\(z^{(2)}\) planes. In (a), \(\ell ^{+}_i\) and \(\ell ^-_i\), \(i = 1, 2\), that represent the lines in (24) and (25) have partitioned the \(a^{(1)}\)-\(z^{(1)}\) plane into nine regions. In (b), \(\ell ^{+}_i\) and \(\ell ^-_i\), represent the lines in (26) and (27), ave partitioned the \(a^{(2)}\)-\(z^{(2)}\) plane into nine regions, where \(1 \le i \le 8\)

The above discussion demonstrates the following proposition.

Proposition A.1

Let \(\mathcal {P}_8 = \{(a^{(1)}, a^{(1)}_r, a^{(2)}, a^{(2)}_r, b^{(2)}, b^{(2)}_r, z^{(1)}, z^{(2)})\}\) be the parameter space of (23). There is a positive integer K and unique set of open subregions \(\{P_k\}_{k=1}^K\) satisfying

  1. (1)

    \(\mathcal {P}_8 = \bigcup _{k=1}^K \overline{P}_k\), where \(\overline{U}\) refers to the closure of U;

  2. (2)

    \(P_i \bigcap P_j = \varnothing \) if \(i \ne j\);

  3. (3)

    Templates \(\mathbb {T}, \mathbb {T}^{\prime } \in P_k\) for some k if and only if \(\mathbf {Y}_{\mathbb {T}} = \mathbf {Y}_{\mathbb {T}^{\prime }}\).

Example A.2

Suppose \(a^{(1)}_r < 0\) and \(a^{(2)}_r, b^{(2)}, b^{(2)}_r\) are all positive. Moreover, choose \(a^{(2)}_r, b^{(2)}, b^{(2)}_r\) such that

$$\begin{aligned} a^{(2)}_r > b^{(2)} > b^{(2)}_r \quad \text {and} \quad a^{(2)}_r > b^{(2)} + b^{(2)}_r. \end{aligned}$$

For instance, \(a^{(1)}_r = -1, a^{(2)}_r = 6, b^{(2)} = 3\), and \(b^{(2)}_r = 2\). Then the position of each line is settled. We number the partitions of \(a^{(1)}\)-\(z^{(1)}\) and \(a^{(2)}\)-\(z^{(2)}\) planes by a pair \([m_{\ell }, n_{\ell }]\), \(\ell = 1, 2\), where \(m_{\ell }, n_{\ell }\) illustrate how many inequalities

$$\begin{aligned} a^{(\ell )} + z^{(\ell )} - 1 > - \xi _{\ell }(\cdot ) \quad \text {and} \quad a^{(\ell )} - z^{(\ell )} - 1 > \xi _{\ell }(\cdot ) \end{aligned}$$

are satisfied, respectively. Thus \(0 \le m_1, n_1 \le 2\) and \(0 \le m_2, n_2 \le 8\). Pick \([m_1, n_1] = [1,2]\) and \([m_2, n_2] = [6,4]\), for instance, \(a^{(1)} = 2, z^{(1)} = -0.3, a^{(2)} = 4\), and \(z^{(2)} = 2.5\). It is easy to check that the basic set of admissible local patterns is

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Ban, JC., Chang, CH. Solution Structure of Multi-layer Neural Networks with Initial Condition. J Dyn Diff Equat 28, 69–92 (2016). https://doi.org/10.1007/s10884-015-9471-9

Download citation

  • Received:

  • Revised:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1007/s10884-015-9471-9

Keywords

Mathematics Subject Classification

Navigation