Parallel information processing by a reservoir computing system based on a VCSEL subject to double optical feedback and optical injection

XiangSheng Tan; YuShuang Hou; YuShuang Hou; ZhengMao Wu; ZhengMao Wu; GuangQiong Xia

doi:10.1364/OE.27.026070

1. Introduction

Nonlinear systems with one or more delay feedback have been attracting considerable attention due to their numerous applications in science and engineering [1–6]. The simplest frame of a delay system only contains a nonlinear device and a delay loop. Recently, such a simple frame provides an exciting prospect for implementing reservoir computing (RC), where RC is a newly-proposed and brain-inspired computational conception [7–9]. In 2011, a RC system based on a nonlinear node and a delay feedback loop (named as delay-based RC) was firstly electrically implemented to process time-dependent signals, and its performance can be comparable to those based on complex network structures [10]. Since then, some related hardware implementations have been reported, such as optoelectronic systems [11–14], all-optical systems [15–17], and laser dynamical systems [18–27]. Particularly, semiconductor lasers (SLs) with time-delay feedback are very promising for high-speed implementation of RC due to their high relaxation oscillation frequency [28]. Brunner et al. experimentally demonstrated a RC system based on an SL subject to both optical feedback and optical injection for performing a time series prediction task, and the lowest prediction error is about 10% for a data rate beyond 1 GByte/s [18]. Uchida et al. theoretically and experimentally investigated the influences of the mask signal format on the performances of a RC based on an SL subject to both optical feedback and optical injection, and the results showed that a better prediction performance can be obtained by adopting a chaotic signal as the mask signal [22,23]. Recently, our group proposed a RC system based on an SL under double optical feedback and optical injection, and the performance of a Santa-Fe time series prediction task can be further improved [26].

In the schemes mentioned above, the used lasers are almost edge-emitting semiconductor lasers (EESLs). Compared with EESLs, vertical-cavity surface-emitting lasers (VCSELs) possess some unique virtues such as small size, low threshold current, single-longitudinal-mode emission, and easiness to photonic integration [29–31]. In particular, VCSELs are capable of emitting two orthogonal polarization components (X-PC and Y-PC) under certain conditions [32]. In recent, based on a VCSEL under an optical feedback and an optical injection, Vatin et al. proposed a RC system, in which the data to be processed was injected into a VCSEL along the main lasing polarization axis. The simulated results demonstrated that, under the case of dual polarization modes co-existing in a VCSEL, the RC system can yield a better performance for processing a single task [29].

In principle, for a VCSEL-based RC system, if the VCSEL can simultaneously output X-PC and Y-PC, each PC can be used to process a task and then the RC system can realize the parallel processing of two independent tasks via the nonlinear dynamical responses of X-PC and Y-PC, which may possess great prospects in today’s big data age. Based on this consideration, in this paper, we propose a RC system based on a VCSEL simultaneously subject to double optical feedback and optical injection to deal with a Santa-Fe time series prediction task and a signal classification task in parallel by using two PCs in the VCSEL. Considering that two different feedback frames (polarization-preserved optical feedback (PP-OF) or polarization-rotated optical feedback (PR-OF)) can be adopted in two feedback loops, there exist four feedback combination cases. For the RC system under four different feedback combinations, the parallel processing ability is examined numerically and compared with each other, and the influences of some typical operation parameters on the RC performance have also been investigated.

2. System model

Figure 1 depicts a schematic diagram of our proposed RC system for processing two independent tasks in parallel, which is based on a VCSEL subject to double optical feedback and optical injection. This system is composed of three main parts: an input layer, a reservoir and an output layer. In the reservoir, a VCSEL subject to double optical feedback supplied by two feedback loops is utilized as a nonlinear node. For convenience, the feedback loop with a shorter length is named as feedback loop 1, and the other loop is named as feedback loop 2. Correspondingly, the delay feedback times for loop 1 and loop 2 are denoted as τ₁ and τ₂, respectively. Two variable optical attenuators (VOA1 and VOA2) and two polarization controllers (PC1 and PC2) are used to control the feedback strengths and the polarization directions of the feedback beams in loop 1 and loop 2, respectively. Here, we define a combination of a polarization-preserved optical feedback (PP-OF) adopted in loop 1 and a polarization-rotated optical feedback (PR-OF) adopted in loop 2 as the case of PP-PR-OF. Accordingly, the other three combination cases of PP-PP-OF, PR-PP-OF and PR-PR-OF correspond to adopting PP-OFs in both two loops, PR-OF in loop 1 and PP-OF in loop 2, PR-OFs in both two loops, respectively. In the input layer, input data 1 is firstly sampled and denoted as U₁(n), where n is a discrete time index and then U₁(n) is held within a period of T. Next, a random sequence with a period of T is applied as a mask signal and multiplied by U₁(n)_. After scaled by a scaling factor γ, the masked input signal is denoted as S₁(t). Input data 2 experiences a similar process and is transferred into S₂(t). S₁(t) and S₂(t) are utilized to respectively modulate the phase of the optical field output from SL1 and SL2 via two phase modulators (PM1 and PM2). The polarization directions of the two modulated optical fields are adjusted to match with X-polarization component (X-PC) and Y-polarization component (Y-PC) of the VCSEL, respectively, and then injected into the VCSEL. In the output layer, via a polarization beam splitter (PBS), X-PC and Y-PC are split. The output of X-PC (|E_x|²) and Y-PC (|E_y|²) is respectively extracted at an interval of θ and interpreted as a state of a virtual node, thus the number N of virtual nodes along loop 1 satisfies N = τ₁/θ. Here, we employ a synchronization scheme, i. e. T = τ₁ [20,21]. The states of N virtual nodes are finally weighted and linearly summed up to obtain the task output. Through minimizing the mean-square error between the target function and the RC output via a linear least-square method, the weights can be optimized [12,22].

Fig. 1 Schematic diagram of a RC system based on a VCSEL subject to double optical feedback and optical injection for processing two independent tasks in parallel. OC: optical circulator; VOA: variable optical attenuator; PC: polarization controller; FC: fiber coupler; PFC: polarization fiber coupler; PBS: polarization beam splitter; PM: phase modulator; SL: semiconductor laser.

Download Full Size | PDF

Based on the spin-flip model (SFM), nonlinear dynamics of a VCSEL subject to double optical feedback and optical injection can be described by [29,33]:

\begin{array}{l} \frac{d E_{x}}{d t} = κ (1 + i α) (N E_{x} - E_{x} + i n E_{y}) - (γ_{a} + i γ_{p}) E_{x} \\ + k_{i n j} ε_{1} (t) + η_{1} Φ_{1 x} (t) + η_{2} Φ_{2 x} (t) + F_{x} (t), \end{array}

\begin{array}{l} \frac{d E_{y}}{d t} = κ (1 + i α) (N E_{y} - E_{y} - i n E_{x}) + (γ_{a} + i γ_{p}) E_{y} \\ + k_{i n j} ε_{2} (t) + η_{1} Φ_{1 y} (t) + η_{2} Φ_{2 y} (t) + F_{y} (t), \end{array}

\frac{d N}{d t} = - γ_{e} [N - μ + N ({| E_{x} |}^{2} + {| E_{y} |}^{2}) + i n (E_{y} E_{x}^{*} - E_{x} E_{y}^{*})],

\frac{d n}{d t} = - γ_{s} n - γ_{e} [n ({| E_{x} |}^{2} + {| E_{y} |}^{2}) + i N (E_{y} E_{x}^{*} - E_{x} E_{y}^{*})],

where subscripts x and y represent X-PC and Y-PC of the VCSEL, respectively. E is the slowly varied complex amplitude of the electric field, N accounts for the total carrier inversion between conduction and valence bands, n is the difference between carrier inversions for the spin-up and spin-down radiation channels, κ accounts for the decay rate of field, α describes the linewidth enhancement factor, γ_e stands for the decay rate of total carrier population, γ_s expresses the spin-flip rate, γ_a and γ_p denote the linear anisotropies representing dichroism and birefringence, respectively. μ is the normalized injection current (μ takes the value 1 at threshold).

The third terms on the right of Eqs. (1) and (2) describe the injection fields containing two independent input signals required to be processed via the dynamical responses of X-PC and Y-PC, respectively. k_inj is the injection strength, and the injected complex electric fields ε₁ and ε₂ are represented as [22]:

ε_{1, 2} (t) = \sqrt{P_{i n j}} e^{i π S_{1, 2} (t)} e^{i 2 π Δ f_{1, 2} t},

where P_inj is the power of CW output of SL. ∆f₁ (∆f₂) is the frequency detuning between the lasing frequency of SL1 (SL2) and the central frequency of the VCSEL. S_1,2 stands for masked input data of two tasks and is written as:

S_{1, 2} (t) = γ \times M a s k_{1, 2} (t) \times U_{1, 2} (n),

where γ is the input scaling factor, Mask_1,2(t) and U_1,2(n) express the mask signals and the sampled and held input signals of two independent tasks, respectively.

The fourth and fifth terms on the right of Eqs. (1) and (2) describe the effects of optical feedback, where η₁ (η₂) is the feedback strength for the loop 1 (loop 2). For four different feedback combination cases, the feedback electric fields are described as follows:

PP - PP - OF : Φ_{1 x, 1 y} (t) = E_{x, y} (t - τ_{1}) e^{- i ω_{0} τ_{1}}, Φ_{2 x, 2 y} (t) = E_{x, y} (t - τ_{2}) e^{- i ω_{0} τ_{2}},

PP - PR - OF : Φ_{1 x, 1 y} (t) = E_{x, y} (t - τ_{1}) e^{- i ω_{0} τ_{1}}, Φ_{2 x, 2 y} (t) = E_{y, x} (t - τ_{2}) e^{- i ω_{0} τ_{2}},

PR - PP - OF : Φ_{1 x, 1 y} (t) = E_{y, x} (t - τ_{1}) e^{- i ω_{0} τ_{1}}, Φ_{2 x, 2 y} (t) = E_{x, y} (t - τ_{2}) e^{- i ω_{0} τ_{2}},

PR - PR - OF : Φ_{1 x, 1 y} (t) = E_{y, x} (t - τ_{1}) e^{- i ω_{0} τ_{1}}, Φ_{2 x, 2 y} (t) = E_{y, x} (t - τ_{2}) e^{- i ω_{0} τ_{2}},

where ω₀ = (ω_x + ω_y)/2 is defined as the central frequency of the VCSEL, and ω_x (ω_y) is the angular frequency of X-PC (Y-PC). In this work, an 850 nm VCSEL is considered, and then ω₀ is about 2.22 × 10¹⁵ Hz.

The last terms on the right of Eqs. (1) and (2) represent Langevin noise sources, which can be expressed as

F_{x} (t) = \sqrt{\frac{β_{s p}}{2}} (\sqrt{N + n} ξ_{1} + \sqrt{N - n} ξ_{2}), F_{y} (t) = - i \sqrt{\frac{β_{s p}}{2}} (\sqrt{N + n} ξ_{1} - \sqrt{N - n} ξ_{2}),

where ξ₁ and ξ₂ are the complex Gaussian white noises with unit variance and zero mean, and β_sp is the spontaneous emission coefficient to be set at 10⁻⁶ ns⁻¹ after referring to [33].

3. Results and discussion

Equations (1)-(4) can be numerically solved by employing a fourth-order Runge-Kutta method with a step of 2 ps via MATLAB software. During the simulation, the following parameter values are used [33]: μ = 1.1, α = 3, κ = 300 ns⁻¹, γ_e = 1 ns⁻¹, γ_s = 50 ns⁻¹, γ_a = 0.1 ns⁻¹, γ_p = 10 ns⁻¹, k_inj = 50 ns^-¹. In this work, the mask signals are randomly chosen from a six-level sequence {-1, −0.6, −0.3, 0.3, 0.6, 1} as Mask₁ for a Santa-Fe time series prediction task and Mask₂ for a signal classification task since the mask with six possible values can offer enough variability of responses for computation [14]. η₁ and η₂ are set 10 ns⁻¹ to guarantee that the system operates at a steady state in absence of input signals for the injected power P_inj varied within a discussed region of (0.2, 10) in this work. The delay time τ₁ of loop 1 is fixed at 8 ns, corresponding to a processing rate of 125 Mb/s. The time difference ∆τ ( = τ₂ – τ₁) between two feedback loops is set as a half of the reciprocal of relaxation oscillation frequency ( $f_{r o} = \sqrt{2 κ γ_{e} (μ - 1)} / 2 π$ ) of the VCSEL at free-running. Under the parameters given above, ∆τ = 0.406 ns [26,34].

Here, X-PC and Y-PC is utilized to execute a Santa-Fe time series prediction task (denoted as TASKx) and a signal classification task (denoted as TASKy), respectively. For TASKx, it needs to predict the next step of a Santa-Fe time series recorded experimentally from a far-infrared laser operating at a chaotic state [35]. A Santa-Fe time series data set contains 9000 points, where consecutive 4000 points (the first 75% is used for training and the remaining 25% is for testing) are selected in this simulation. For TASKy, input data is a random concatenation of sine and square waves, each of which is discretized into 12 points. 250 waves (3000 points) are selected for training and 83 waves (1000 points) for testing, with the target value being 0 for a sine wave and 1 for a square wave. For both tasks, the Normalized Mean Square Error (NMSE) between the target value ӯ(n) and the reservoir output y(n) is calculated to measure the performance of the reservoir, which is defined as [22]:

N M S E = \frac{1}{L} \frac{{\sum_{n = 1}^{L} (y (n) - \bar{y} (n))}^{2}}{v a r (y)},

where n is the discrete time index of input data and L is the total number of data in the testing data set, var(y) represents the variance. Typically, a RC system can be considered to perform well when NMSE is less than 0.1 for TASKx and 2 × 10⁻⁴ for TASKy, respectively.

First, we investigate the influence of the number N of virtual nodes on the parallel processing ability of the VCSEL-based RC system. Figure 2 presents the variations of the mean NMSE of ten calculations for TASKx (Fig. 2(a)) and TASKy (Fig. 2(b)) with the number N of virtual nodes under four different feedback combination cases. It can be seen that, for four different cases, the NMSE values for processing two tasks firstly decrease rapidly and then maintain at a relatively stable level with the increase of N. The reason that a less virtual node results in a larger mean NMSE value may be explained as follows. In this work, τ₁ = Nθ is fixed at 8 ns, a smaller N is accompanied by a larger node interval θ. On one hand, a smaller N means a lower dimension of the state space, which makes the training of the system be implemented more difficult and therefore results in a larger NMSE. On the other hand, a large θ may leads to an insufficient coupling between two adjacent nodes. As a result, the joint action of above two factors may cause a relatively bad performance. A relatively stable level of NMSE value maintained under a large N can be explained as follows. As demonstrated in [10], for a fixed N, with the decrease of θ the processing performance is degraded due to the increase of the correlation between two adjacent nodes. However, in this work, Nθ ( = τ₁) is fixed at 8 ns, i.e. a decrease of θ is accompanied by an increase of N. As a result, the degradation originating from the decreasing θ can be compensated through the increase of N, and then the performance of the RC system can maintain at a relatively stable level. Through comparing the four feedback cases, it can be found that the NMSE values for the RC under PP-PP-OF are always lower than those for the RC under the other three cases. The different processing ability of the RC system is owing to the complex coupling between X-PC and Y-PC under different feedback frames. For N ≥ 500, all the NMSE values of TASKx for the RC under the four feedback cases are lower than 0.1, while the NMSE values of TASKy for the RC are lower than 2 × 10⁻⁴ except the case of PP-PR-OF. Additionally, for comparison, the performance of the RC system under PP-PP-OF for processing a single task is also shown as dashed lines with square marks in this figure. Obviously, after adding another task, the system performance is degraded. However, the prediction performance of the parallel processing of the RC system is similar to that of a single prediction task of the system if N ≥ 500.

Fig. 2 Mean NMSE of 10 calculations for (a) TASKx and (b) TASKy as a function of the number N of virtual nodes with ∆f_{1, 2} = 0 GHz, P_inj = 1 and γ = 1. The solid lines with solid marks correspond to the parallel processing of two tasks by the RC system under four feedback combination cases, and the dashed lines with hollow square marks corresponds to a single task processing by the RC system under PP-PP-OF.

Download Full Size | PDF

In order to intuitively observe the parallel processing ability of this proposed VCSEL-based RC system, the parallel processing results of TASKx and TASKy under PP-PP-OF with N = 500 are taken as an example and shown in Fig. 3. It can be seen that the output values of this proposed RC system are almost consistent with the target values for both tasks, and the NMSEs are 0.0364 for TASKx and 1.26 × 10⁻⁴ for TASKy, respectively.

Fig. 3 Performance illustrations of TASKx (the left column) and TASKy (the right column) under PP-PP-OF with N = 500, ∆f_{1, 2} = 0 GHz, P_inj = 1, and γ = 1. (a) A sample of target chaotic Santa-Fe time series data ӯ(n); (b) Predicted output y(n) of the proposed VCSEL-based RC system; (c) Error (y(n)- ӯ(n)) between the target value and the reservoir output; (d) A set of input data composed of random concatenation of sine and square waves; (e) Target values, where 0 and 1 correspond sine and square waves, respectively; (f) Classification output of the proposed VCSEL-based RC system.

Download Full Size | PDF

Second, we investigate the NMSEs of two different tasks processed by this RC system when the frequency detuning ∆f is changed. For simplification, we set ∆f₁ = ∆f₂ = ∆f. Moreover, considering that the performance of the proposed RC system changes very slightly when N ϵ [500, 1000], in the following discussion we fix N = 500. Figure 4 displays the dependence of the NMSEs on ∆f for two tasks under different feedback cases. For TASKx as shown in Fig. 4(a), it can be seen that, within the frequency detuning region [-32 GHz, 16 GHz], almost all the NMSE values for the RC system under the three cases of PP-PP-OF, PP-PR-OF and PR-PP-OF are lower than 0.1, and the NMSE values for the RC under PP-PP-OF are almost the lowest. However, few NMSE values for the RC under PR-PR-OF are less than 0.1. As for TASKy (Fig. 4(b)), three feedback frames including PP-PP-OF, PR-PP-OF and PR-PR-OF possess similar NMSEs lower than 2 × 10⁻⁴ within −24 GHz ≤ ∆f ≤ 20 GHz. Comparing Fig. 4(a) with Fig. 4(b), it can be seen that there exist some differences between TASKx and TASKy, which may originate from different properties of the two tasks and the complex coupling between X-PC and Y-PC. Generally speaking, the RC system adopting PP-PP-OF would be helpful for yielding a well parallel processing performance. In particular, for ∆f = −12 GHz, the values of NMSE for both two tasks arrive at their minima.

Fig. 4 Mean NMSE of 10 calculations for (a) TASKx and (b) TASKy as a function of the frequency detuning ∆f under cases of PP-PP-OF (blue lines with square marks), PP-PR-OF (green lines with triangle marks), PR-PP-OF (black lines with cross marks), or PR-PR-OF (red lines with circle marks) with N = 500, P_inj = 1, and γ = 1.

Download Full Size | PDF

Next, the influence of the injection power P_inj on the parallel processing ability of the proposed RC system is examined. Figure 5 shows the NMSEs for the RC under PP-PP-OF with ∆f = −12 GHz and varying injection powers within (0.2, 10), where the error bars are the statistical results of 10 calculations. For processing TASKx (as shown in Fig. 5(a)), it can be obviously observed that, with the increase of P_inj, the NMSE values firstly decrease sharply, after passing through a minimum, and then increase slowly. As a result, there exists an optimal injected power, which is coincided with that reported in [29]. For processing TASKy (as shown in Fig. 5(b)), a similar variation trend with Fig. 5(a) can be observed. Therefore, a moderate injected power is more suitable for achieving a well RC performance.

Fig. 5 Mean NMSE of 10 calculations for processing (a) TASKx and (b) TASKy as a function of the injected power P_inj for the RC under PP-PP-OF with N = 500, ∆f = −12 GHz and γ = 1, where error bars are the statistical results of 10 calculations.

Download Full Size | PDF

Previous researches have demonstrated that the computational performance of a RC system is also dependent on the scaling factor γ [22,36]. Finally, we calculate the evolution of NMSEs for processing TASKx and TASKy in the parameter space of P_inj and γ, as shown in Fig. 6. Here, we take the RC under PP-PP-OF as an example. In this diagram, different colors correspond to different NMSE values, and the regions surrounded by white dashed lines satisfy NMSE ≤ 0.1 in Fig. 6(a) and NMSE ≤ 2 × 10⁻⁴ in Fig. 6(b), in which the RC system presents a well processing ability. Obviously, the parameter region with NMSE ≤ 2 × 10⁻⁴ for implementing TASKy is wider than that with NMSE ≤ 0.1 for implementing TASKx. In order to achieve a well parallel processing ability, P_inj and γ should be located at a overlapped region of the two regions surrounded by the white lines in Fig. 6(a) and 6(b). The minimal NMSE value is 0.0289 for TASKx and 2.78 × 10⁻⁵ for TASKy.

Fig. 6 Evolution maps of the NMSE values in a P_inj – γ parameter space for the RC under PP-PP-OF to implement (a) TASKx and (b) TASKy with N = 500 and ∆f = −12 GHz.

Download Full Size | PDF

4. Conclusions

In summary, a RC system based on a vertical-cavity surface-emitting laser (VCSEL) subject to double optical feedback and optical injection is proposed and numerically investigated. For this system, the dynamical responses of two polarization components (named as X-PC and Y-PC) of a VCSEL are utilized to process a Santa-Fe time series prediction task and a signal classification task, respectively. Taking into account four different feedback combination frames adopted in two feedback loops including PP-PP-OF, PP-PR-OF, PR-PP-OF and PR-PR-OF, we simulate the dependence of the parallel processing ability of the RC on certain typical operation parameters such as the number N of virtual nodes, the frequency detuning ∆f, the injection power P_inj_, and the input scaling factor γ. The results indicate that, for these parameters varied within a relatively broad region, this VCSEL-based RC system can well process two independent tasks in parallel. Through comparing the processing performance of the RC under the four feedback combination cases, the best parallel processing ability can be achieved under a PP-PP-OF. As a result, for optimized parameters the RC system can process a Santa-Fe time series prediction task with a minimal error of 0.0289 and a signal classification task with a minimal error of 2.78 × 10⁻⁵, respectively.,

Funding

National Natural Science Foundation of China (NSFC) (61575163, 61775184, 61875167).

References

1. T. Erneux, Applied Delayed Differential Equations (Springer, 2009).

2. R. Boucekkine, D. D. L. Croix, and O. Licandro, “Modelling vintage structures with DDEs: Principles and applications,” Math. Popul. Stud. 11(3), 151–179 (2004). [CrossRef]

3. R. J. Henry, Z. N. Masoud, A. H. Nayfeh, and D. T. Mook, “Cargo pendulation reduction on ship-mounted cranes via boom-lu angle actuation,” J. Vib. Control 7(8), 1253–1264 (2001). [CrossRef]

4. Z. N. Masoud, A. H. Nayfeh, and A. Al-Mousa, “Delayed position feedback controller for the reduction of payload pendulations of rotary cranes,” J. Vib. Control 9(1), 257–277 (2003). [CrossRef]

5. N. Q. Li, R. M. Nguimdo, A. Locquet, and D. S. Citrin, “Enhancing optical-feedback-induced chaotic dynamics in semiconductor ring lasers via optical injection,” Nonlinear Dyn. 92(2), 315–324 (2018). [CrossRef]

6. J. Ohtsubo, Semiconductor Lasers: Stability, Instability and Chaos (Springer, 2013).

7. D. Brunner, B. Penkovsky, B. A. Marquez, M. Jacquot, I. Fischer, and L. Larger, “Tutorial: Photonic neural networks in delay systems,” J. Appl. Phys. 124(15), 152004 (2018). [CrossRef]

8. H. Jaeger and H. Haas, “Harnessing nonlinearity: predicting chaotic systems and saving energy in wireless communication,” Science 304(5667), 78–80 (2004). [CrossRef] [PubMed]

9. D. Verstraeten, B. Schrauwen, M. D’Haene, and D. Stroobandt, “An experimental unification of reservoir computing methods,” Neural Netw. 20(3), 391–403 (2007). [CrossRef] [PubMed]

10. L. Appeltant, M. C. Soriano, G. Van der Sande, J. Danckaert, S. Massar, J. Dambre, B. Schrauwen, C. R. Mirasso, and I. Fischer, “Information processing using a single dynamical node as complex system,” Nat. Commun. 2(1), 468 (2011). [CrossRef] [PubMed]

11. L. Larger, M. C. Soriano, D. Brunner, L. Appeltant, J. M. Gutierrez, L. Pesquera, C. R. Mirasso, and I. Fischer, “Photonic information processing beyond Turing: an optoelectronic implementation of reservoir computing,” Opt. Express 20(3), 3241–3249 (2012). [CrossRef] [PubMed]

12. Y. Paquot, F. Duport, A. Smerieri, J. Dambre, B. Schrauwen, M. Haelterman, and S. Massar, “Optoelectronic reservoir computing,” Sci. Rep. 2(1), 287 (2012). [CrossRef] [PubMed]

13. R. Martinenghi, S. Rybalko, M. Jacquot, Y. K. Chembo, and L. Larger, “Photonic nonlinear transient computing with multiple-delay wavelength dynamics,” Phys. Rev. Lett. 108(24), 244101 (2012). [CrossRef] [PubMed]

14. M. C. Soriano, S. Ortín, D. Brunner, L. Larger, C. R. Mirasso, I. Fischer, and L. Pesquera, “Optoelectronic reservoir computing: tackling noise-induced performance degradation,” Opt. Express 21(1), 12–20 (2013). [CrossRef] [PubMed]

15. F. Duport, B. Schneider, A. Smerieri, M. Haelterman, and S. Massar, “All-optical reservoir computing,” Opt. Express 20(20), 22783–22795 (2012). [CrossRef] [PubMed]

16. A. Dejonckheere, F. Duport, A. Smerieri, L. Fang, J. L. Oudar, M. Haelterman, and S. Massar, “All-optical reservoir computer based on saturation of absorption,” Opt. Express 22(9), 10868–10881 (2014). [CrossRef] [PubMed]

17. Q. Vinckier, F. Duport, A. Smerieri, K. Vandoorne, P. Bienstman, M. Haelterman, and S. Massar, “High performance photonic reservoir computer based on a coherently driven passive cavity,” Optica 2(5), 438–446 (2015). [CrossRef]

18. D. Brunner, M. C. Soriano, C. R. Mirasso, and I. Fischer, “Parallel photonic information processing at gigabyte per second data rates using transient states,” Nat. Commun. 4(1), 1364 (2013). [CrossRef] [PubMed]

19. K. Hicke, M. A. Escalona-Morán, D. Brunner, M. C. Soriano, I. Fischer, and C. R. Mirasso, “Information processing using transient dynamics of semiconductor lasers subject to delayed feedback,” IEEE J. Sel. Top. Quantum Electron. 19(4), 1501610 (2013). [CrossRef]

20. R. M. Nguimdo, G. Verschaffelt, J. Danckaert, and G. Van der Sande, “Fast photonic information processing using semiconductor lasers with delayed optical feedback: role of phase dynamics,” Opt. Express 22(7), 8672–8686 (2014). [CrossRef] [PubMed]

21. R. M. Nguimdo, G. Verschaffelt, J. Danckaert, and G. Van der Sande, “Reducing the phase sensitivity of laser-based optical reservoir computing systems,” Opt. Express 24(2), 1238–1252 (2016). [CrossRef] [PubMed]

22. J. Nakayama, K. Kanno, and A. Uchida, “Laser dynamical reservoir computing with consistency: an approach of a chaos mask signal,” Opt. Express 24(8), 8679–8692 (2016). [CrossRef] [PubMed]

23. Y. Kuriki, J. Nakayama, K. Takano, and A. Uchida, “Impact of input mask signals on delay-based photonic reservoir computing with semiconductor lasers,” Opt. Express 26(5), 5777–5788 (2018). [CrossRef] [PubMed]

24. R. M. Nguimdo, G. Verschaffelt, J. Danckaert, and G. Van der Sande, “Simultaneous computation of two independent tasks using reservoir computing based on a single photonic nonlinear node with optical feedback,” IEEE Trans. Neural Netw. Learn. Syst. 26(12), 3301–3307 (2015). [CrossRef] [PubMed]

25. R. M. Nguimdo, E. Lacot, O. Jacquin, O. Hugon, G. Van der Sande, and H. Guillet de Chatellus, “Prediction performance of reservoir computing systems based on a diode-pumped erbium-doped microchip laser subject to optical feedback,” Opt. Lett. 42(3), 375–378 (2017). [CrossRef] [PubMed]

26. Y. Hou, G. Xia, W. Yang, D. Wang, E. Jayaprasath, Z. Jiang, C. Hu, and Z. Wu, “Prediction performance of reservoir computing system based on a semiconductor laser subject to double optical feedback and optical injection,” Opt. Express 26(8), 10211–10219 (2018). [CrossRef] [PubMed]

27. R. M. Nguimdo and T. Erneux, “Enhanced performances of a photonic reservoir computer based on a single delayed quantum cascade laser,” Opt. Lett. 44(1), 49–52 (2019). [CrossRef] [PubMed]

28. A. Uchida, Optical Communication with Chaotic Lasers: Applications of Nonlinear Dynamics and Synchronization (Wiley-VCH, 2012).

29. J. Vatin, D. Rontani, and M. Sciamanna, “Enhanced performance of a reservoir computer using polarization dynamics in VCSELs,” Opt. Lett. 43(18), 4497–4500 (2018). [CrossRef] [PubMed]

30. S. Y. Xiang, W. Pan, B. Luo, L. Yan, X. Zou, N. Jiang, L. Yang, and H. Zhu, “Unpredictability-enhanced chaotic vertical-cavity surface-emitting lasers with variable-polarization optical feedback,” J. Lightwave Technol. 29(14), 2173–2179 (2011). [CrossRef]

31. F. Koyama, “Recent advances of VCSEL photonics,” J. Lightwave Technol. 24(12), 4502–4513 (2006). [CrossRef]

32. J. Martin-Regalado, F. Prati, M. San Miguel, and N. B. Abraham, “Polarization properties of vertical-cavity surface-emitting lasers,” IEEE J. Quantum Electron. 33(5), 765–783 (1997). [CrossRef]

33. I. Gatare, M. Sciamanna, A. Locquet, and K. Panajotov, “Influence of polarization mode competition on the synchronization of two unidirectionally coupled vertical-cavity surface-emitting lasers,” Opt. Lett. 32(12), 1629–1631 (2007). [CrossRef] [PubMed]

34. J. G. Wu, G. Q. Xia, and Z. M. Wu, “Suppression of time delay signatures of chaotic output in a semiconductor laser with double optical feedback,” Opt. Express 17(22), 20124–20133 (2009). [CrossRef] [PubMed]

35. A. S. Weigend and N. A. Gershenfeld, “Time series prediction: forecasting the future and understanding the past,” http://www-psych.stanford.edu/~andreas/Time-Series/SantaFe.html (1993).

36. L. Appeltant, “Reservoir computing based on delay-dynamical systems,” http://www.tdx.cat/handle/10803/84144 (2012).

Parallel information processing by a reservoir computing system based on a VCSEL subject to double optical feedback and optical injection

Abstract

1. Introduction

2. System model

3. Results and discussion

4. Conclusions

Funding

References

Cited By

Figures (6)

Equations (12)

Optics Express