Semi-supervised deep learning by metric embedding
Authors: Elad Hoffer, Nir Ailon
ICLR 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | All experiments were conducted using the Torch7 framework by Collobert et al. (2011). Code reproducing these results will by available at https://github.com/eladhoffer/SemiSupContrast. For every experiment we chose a small random subset of examples, with a balanced number from each class and denoted by XL. The remaining training images are used without their labeled to form XU. Finally, we test our final accuracy with a disjoint set of examples Xtest. No data augmentation was applied to the training sets. |
| Researcher Affiliation | Academia | Elad Hoffer Technion Israel Institute of Technology Haifa, Israel ehoffer@tx.technion.ac.ilNir Ailon Technion Israel Institute of Technology Haifa, Israel nailon@cs.technion.ac.il |
| Pseudocode | No | The paper describes mathematical formulations and processes but does not include any pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code reproducing these results will by available at https://github.com/eladhoffer/SemiSupContrast. |
| Open Datasets | Yes | The MNIST database of handwritten digits introduced by Le Cun et al. (1998) is one of the most studied dataset benchmark for image classification. The dataset contains 60,000 examples of handwritten digits from 0 to 9 for training and 10,000 additional examples for testing, where each sample is a 28 x 28 pixel gray level image. |
| Dataset Splits | Yes | The choices for λ and k-NN parameters were made using a validation set. We did not found any substantial difference between the values we explored, so they were usually left as the default value for simplicity. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. It only mentions the software framework: "All experiments were conducted using the Torch7 framework by Collobert et al. (2011)." |
| Software Dependencies | Yes | All experiments were conducted using the Torch7 framework by Collobert et al. (2011). |
| Experiment Setup | Yes | A batch-size of b = 32 was used for all experiments, totaling a sampled set of 2 b + |C| examples for each iteration, where |C| = 10 for both datasets. We used 6 as optimization criterion, where λL = λU = 1. Optimization was done using the Accelerated-gradient method by Nesterov (1983) with an initial learning rate of lr0 = 0.1 which was decreased by a factor of 10 after every 30 epochs. Both datasets were trained on for a total of 90 epochs. Final test accuracy results was achieved by using a k-NN classifier with best results out of k = {1, 3, 5}. |