Representation Learning via Adversarially-Contrastive Optimal Transport
Authors: Anoop Cherian, Shuchin Aeron
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present experiments demonstrating the effectiveness of our approach for representation learning. To this end, we use two standard action recognition datasets, namely (i) small-scale JHMDB (Jhuang et al., 2013) and (ii) the larger HMDB dataset (Kuehne et al., 2011). |
| Researcher Affiliation | Collaboration | 1Mitsubishi Electric Research Labs, Cambridge, MA. 2Tufts University, Medford, MA. |
| Pseudocode | No | The paper describes algorithmic steps in prose and mathematical formulations, but no explicit pseudocode or algorithm blocks are provided. |
| Open Source Code | No | We will be making our code publicly available at https://www.merl.com/research/license/. |
| Open Datasets | Yes | JHMDB dataset: consists of 928 video sequences, each sequence about 15-40 frames long. There are 21 actions defined on the clips... HMDB Dataset: is a super-set of the JHMDB dataset and consists of about 6700 video clips... Qualitative Visualizations: To gain insights into the kind of perturbations our adversarial network generates, we trained this sub-module on the CIFAR10 dataset. |
| Dataset Splits | Yes | The results are based on the split-1 of the respective datasets. |
| Hardware Specification | No | No specific hardware details such as GPU/CPU models, memory, or cloud instance specifications are mentioned for running experiments. |
| Software Dependencies | No | Our entire implementation is in Py Torch. For our adversarial module, we modified the public WGAN code associated with (Arjovsky et al., 2017). We used Py Man Opt as our Riemannian optimization framework. |
| Experiment Setup | Yes | Our entire implementation is in Py Torch. For our adversarial module, we modified the public WGAN code associated with (Arjovsky et al., 2017). We used a noise variance σ = 0.01, which resulted in an average classifier fooling rate of 60% on the training set on both the datasets. See the Appendix for more experiments in this regard. Further, we used λ1 = 0.1 and λ2 = 1 in (9). We used Py Man Opt as our Riemannian optimization framework4. As for the regularization constants on the distortion and ordering constraints in (10), we set β1 = 1 and β2 = 10, and we used η = 0.01 for the temporal margin. |