Unfolding recurrence by Green’s functions for optimized reservoir computing
Authors: Sandra Nestler, Christian Keup, David Dahmen, Matthieu Gilson, Holger Rauhut, Moritz Helias
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conclude the analysis with an application to a univariate temporal classification dataset. This serves as a proof-of-concept to demonstrate the effects of the optimization on a real-world problem and can be regarded as a check that real data do not generally contain structural obstacles that were not covered in the theoretical considerations. To raise the method from the proof-of-concept level, the performance should be systematically checked on a broader set of problems as done for state of the art time series classifiers [Bagnall et al., 2017, Wang et al., 2017], which we leave for future work. The examined dataset is ECG5000, which is publicly available at the UCR Time Series Classification archive [Chen et al., 2015], containing 5000 electrocardiograms of single heartbeat recordings. The classes separate between five categories of healthy and diseased heartbeats. For a binary classification, we use only samples from the two largest classes, so that we obtained a training set consisting of 354 samples and a testing set of 4332 samples. All stimuli were shifted and scaled to provide classes with means µ with µ = 1; higher order cumulants changed accordingly. This scaling of inputs is only performed for conceptual clarity, allowing identical network parameters as in the previous task. Likewise, one could adapt the value of α according to the stimulus strength. Furthermore, for maximal performance, a trained threshold can replace the centering of data. As a measure of linear separability, we relate the difference of the class means to the covariance in the direction of separation. This yields a ratio µ 2/ µT ψµ = 2.6, which is much higher than for the artificial stimuli analyzed in figure 3, where the corresponding measure ranges between 0.19 and 0.98. All results presented here use the same parameters as in figure 2 and figure 3. The summary of the results in table 1, which contains averaged results over 20 different initializations of the recurrent connectivity, makes evident that a maximized soft margin is accompanied by increased accuracies. The optimized input projections outperformed all randomly chosen ones both with respect to soft margin and accuracy. |
| Researcher Affiliation | Academia | 1Institute of Neuroscience and Medicine (INM-6), Jülich Research Centre, Jülich, Germany 2Mathematics of Information Processing, RWTH Aachen University, Aachen, Germany 3Department of Physics, RWTH Aachen University, Aachen, Germany 4Universitat Pompeu Fabra, Barcelona, Spain 5RWTH Aachen University, Aachen, Germany |
| Pseudocode | No | More details and pseudocode can be found in the supplementary material (section A.3). The provided text does not contain pseudocode or a clearly labeled algorithm block. |
| Open Source Code | No | The paper does not provide an explicit statement or link for open-source code for the methodology described. |
| Open Datasets | Yes | The examined dataset is ECG5000, which is publicly available at the UCR Time Series Classification archive [Chen et al., 2015] |
| Dataset Splits | No | For a binary classification, we use only samples from the two largest classes, so that we obtained a training set consisting of 354 samples and a testing set of 4332 samples. The paper specifies training and testing sets but does not explicitly mention a validation set or its split. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for experiments. |
| Software Dependencies | No | The paper does not mention specific software dependencies with version numbers. |
| Experiment Setup | Yes | All results presented here use the same parameters as in figure 2 and figure 3. Figure 2 caption: N = 100 neurons, τ = 0.25 and fixed connectivity W (Wij N(0, g2/N), g = 0.9); Figure 3 caption: α = 0.05 for the non-linear system. |