Sample and Predict Your Latent: Modality-free Sequential Disentanglement via Contrastive Estimation
Authors: Ilan Naiman, Nimrod Berman, Omri Azencot
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our approach on video, audio, and time series benchmarks. Our method presents state-of-the-art results in comparison to existing techniques. |
| Researcher Affiliation | Academia | 1Department of Computer Science, Ben Gurion University of the Negev, Beer-Sheva, Israel. |
| Pseudocode | Yes | Algorithm 1 Static predictive sampling trick |
| Open Source Code | Yes | The code is available at Git Hub. |
| Open Datasets | Yes | Sprites. A dataset introduced by (Reed et al., 2015)... MUG. A Facial expression dataset created by (Aifanti et al., 2010)... TIMIT. A dataset introduced by (Garofolo et al., 1992)... Jester. A dataset introduced by (Materzynska et al., 2019)... Letters. The Letters dataset (Ibrahim et al., 2019)... Physionet. The Physionet ICU Dataset (Goldberger et al., 2000)... Air Quality. The UCI Beijing Multi-site Air Quality dataset (Zhang et al., 2017)... |
| Dataset Splits | No | The paper specifies training and testing splits for datasets, and describes splitting the test set for downstream tasks, but does not explicitly detail a separate validation set split used for tuning the main model during training. |
| Hardware Specification | No | The paper does not specify any hardware details like GPU models, CPU types, or memory used for running the experiments. |
| Software Dependencies | No | All the models have been implemented using Pytorch (Paszke et al., 2019). While PyTorch is mentioned, a specific version number for PyTorch or other software dependencies like Python or CUDA is not provided. |
| Experiment Setup | Yes | The hyperparameter λ1 is tuned over {1, 2.5, 5, 10}, λ2 is tuned over {1, 3, 5, 7, 9}, and λ4 and λ5 are tuned over {0.1, 0.5, 1, 2.5, 5} while λ3 is fixed to 1. We used Adam optimizer (Kingma & Ba, 2014) with the learning rate chosen from {0.001, 0.0015, 0.002}. The static and dynamic features dimensions are selected from {128, 256} and {32, 64}, respectively. |