Gradual Domain Adaptation via Manifold-Constrained Distributionally Robust Optimization
Authors: seyed amir saberi, Amir Najafi, Amin Behjati, Ala Emrani, Yasaman Zolfimoselo, Shadrooy, Abolfazl Motahari, Babak Khalaj
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We have substantiated our theoretical findings through several experimental results.We further validate our theoretical findings through a series of experiments. |
| Researcher Affiliation | Academia | Department of Electrical Engineering, Department of Computer Engineering, Sharif Center for Information Systems and Data Science, Sharif University of Technology, Tehran, Iran |
| Pseudocode | Yes | Algorithm 1: DRO-based Domain Adaptation (DRODA) |
| Open Source Code | Yes | Our supplemental material includes all our codes, and we cite the data we used for our experiments. |
| Open Datasets | Yes | We implemented this method on the 'Rotating MNIST' dataset, similar to [KML20]. |
| Dataset Splits | Yes | In particular, we sampled 6 batches, each with a size of 4200, without replacement from the MNIST dataset, and labeled these batches as D0, D1, , D4, which represent the datasets obtained from P0, P1, , P4, respectively. The images in dataset Di were then rotated by i 15 degrees, with D0 serving as the source dataset and D4 as the target dataset. We provided the source dataset with labels and left D1, D2, D3, and D4 unlabeled for our algorithm. We then tested the accuracy of θ 0, , θ 3 the outputs of our algorithm at each step on D1, D2, D3, and D4, respectively. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using 'Py Torch' but does not specify its version number or any other software dependencies with specific versions. |
| Experiment Setup | Yes | In our experiments, we employed a two-layer CNN with a 7 7 kernel in the first layer and a 5 5 kernel in the second layer for P. (...) For the classifier family C, we used a three-layer CNN with max pooling and a fully connected layer, applying dropout with a rate of 0.5 in the fully connected layer. A standard Stochastic Gradient Descent (SGD) procedure has been used for the min-max optimization procedure described in (26). |