Unsupervised Feature Extraction by Time-Contrastive Learning and Nonlinear ICA
Authors: Aapo Hyvarinen, Hiroshi Morioka
NeurIPS 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 6 Simulation on artificial data and 7 Experiments on real brain imaging data and Results Figure 2a) shows that after training the feature-MLP by TCL, the MLR achieved higher classification accuracies than chance level |
| Researcher Affiliation | Academia | 1 Department of Computer Science and HIIT University of Helsinki, Finland 2 Gatsby Computational Neuroscience Unit University College London, UK |
| Pseudocode | No | The paper describes steps for TCL in paragraph format, but does not include a structured pseudocode or algorithm block. |
| Open Source Code | No | The paper does not include a link or explicit statement about providing open-source code for the described methodology. |
| Open Datasets | Yes | We used MEG data from an earlier neuroimaging study [25], graciously provided by P. Ramkumar. |
| Dataset Splits | Yes | Classification was performed using a linear support vector machine (SVM) classifier trained on the stimulation modality labels, and its performance was evaluated by a session-average of session-wise one-block-out cross-validation (CV) accuracies. The hyperparameters of the SVM were determined by nested CV without using the test data. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU/GPU models, memory, or cloud instance types) used for running experiments. |
| Software Dependencies | No | The paper mentions methods like back-propagation, Fast ICA [15], MLP, MLR, and SVM, but does not provide specific software names with version numbers. |
| Experiment Setup | Yes | Nonstationary source signals (n = 20, segment length 512), As the activation function in the hidden layers, we used a maxout unit, constructed by taking the maximum across G = 2 affine fully connected weight groups. However, the output layer has absolute value activation units exclusively., To train the MLP, we used back-propagation with a momentum term. To avoid overfitting, we used ℓ2 regularization for the feature-MLP and MLR., segments of equal size, of length 12.5 s or 625 data points (downsampling to 50 Hz), The number of layers took the values L {1, 2, 3, 4}, and the number of nodes of each hidden layer was a function of L so that we always fixed the number of output layer nodes to 10, and increased gradually the number of nodes when going to earlier layer as L = 1 : 10, L = 2 : 20 10, L = 3 : 40 20 10, and L = 4 : 80 40 20 10., We used Re LU s in the middle layers, and adaptive units φ(x) = max(x, ax) exclusively for the output layer, To prevent overfitting, we applied dropout [28] to inputs, and batch normalization [19] to hidden layers. |