Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
Photonic Delay Systems as Machine Learning Implementations
Authors: Michiel Hermans, Miguel C. Soriano, Joni Dambre, Peter Bienstman, Ingo Fischer
JMLR 2015 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform physical experiments that demonstrate that the obtained input encodings work well in reality, and we show that optimized systems perform significantly better than the common Reservoir Computing approach. We tested the use of BPTT for training the input masks both in simulation and in experiment on two benchmark tasks. First, we considered the often-used MNIST written digit recognition data set, where we use the dynamics of the system indirectly. Next, we applied it on the TIMIT phoneme data set. |
| Researcher Affiliation | Academia | OPERA Photonics Group Universit e Libre de Bruxelles Avenue F. Roosevelt 50, 1050 Brussels, Belgium. Instituto de F ısica Interdisciplinar y Sistemas Complejos, IFISC (UIB-CSIC) Campus Universitat de les Illes Balears E-07122 Palma de Mallorca, Spain. ELIS departement Ghent University Sint Pietersnieuwstraat 41, 9000 Ghent, Belgium. INTEC departement Ghent University Sint Pietersnieuwstraat 41, 9000 Ghent, Belgium. |
| Pseudocode | No | The paper describes mathematical update equations and experimental procedures in paragraph form, but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statements about releasing source code for the methodology, nor does it provide links to any code repositories. |
| Open Datasets | Yes | We tested the use of BPTT for training the input masks both in simulation and in experiment on two benchmark tasks. First, we considered the often-used MNIST written digit recognition data set, where we use the dynamics of the system indirectly. Next, we applied it on the TIMIT phoneme data set. John Garofolo, National Institute of Standards, Technology (US, Linguistic Data Consortium, Information Science, Technology Office, United States, and Defense Advanced Research Projects Agency). TIMIT Acoustic-phonetic Continuous Speech Corpus. Linguistic Data Consortium, 1993. |
| Dataset Splits | Yes | After training the input weights, we gathered both physical and simulated data for the 4 experiments as described below, and retrained the output weights to obtain a final score. ... as we only presented the DCMZ with the original 60,000 training examples. Meta-parameter optimization was performed using 10,000 randomly selected examples from the training set. Input masks are trained using 50,000 training iterations, where for each iteration the gradient was determined on 200 randomly sampled sequences of a length of 50 frames. |
| Hardware Specification | No | The paper describes the components of the physical delay system (laser source, Mach-Zehnder modulator, optical fiber, electronic circuit) which is the subject of the study, but does not provide specific details about the computational hardware (e.g., GPU/CPU models, memory) used for running the simulations or training the machine learning models. |
| Software Dependencies | No | The paper does not provide specific software dependencies or their version numbers used for the experiments or simulations. |
| Experiment Setup | Yes | Input masks were trained using 10^6 training iterations, where for each iteration the gradient was determined on 500 randomly sampled digits. For training we used Nesterov momentum (Sutskever et al., 2013), with momentum coefficient 0.9, and a learning rate of 0.01 which linearly decayed to zero over the duration of the training. Output weights are trained using the cross-entropy loss function over 10^6 training iterations, where for each iteration the gradient was determined on 1000 randomly sampled digits. We again used Nesterov momentum, with momentum coefficient 0.9. The learning rate was chosen at 0.002 and linearly decayed to zero. Meta-parameter optimization was performed using 10,000 randomly selected examples from the training set. |