Expected Information Maximization: Using the I-Projection for Mixture Density Estimation
Authors: Philipp Becker, Oleg Arenz, Gerhard Neumann
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We compare to the f-GAN which is the only other method capable of minimizing the I-projection solely based on samples. We use data sampled from randomly generated GMMs with different numbers of components and dimensionalities. |
| Researcher Affiliation | Collaboration | Philipp Becker Autonomous Learning Robots, KIT Bosch Center for Artificial Intelligence Oleg Arenz Intelligent Autonomous Systems, TU Darmstadt Gerhard Neumann Autonomous Learning Robots, KIT Bosch Center for Artificial Intelligence University of T ubingen |
| Pseudocode | Yes | Pseudo-code for EIM for GMMs can be found in algorithm 1 |
| Open Source Code | Yes | Code available at https://github.com/pbecker93/Expected Information Maximization |
| Open Datasets | Yes | We evaluated our approach on data from the Stanford Drone Dataset (Robicquet et al., 2016) and a traffic dataset from the Next Generation Simulation program. |
| Dataset Splits | Yes | 10, 000 Train Samples, 5, 000 Test Samples, 5, 000 Validation samples (for early stopping the density ratio estimator) |
| Hardware Specification | No | The paper does not specify any particular hardware (e.g., CPU, GPU model numbers or specific memory amounts) used for running the experiments. |
| Software Dependencies | No | The paper mentions 'Adam (Kingma & Ba, 2014)' as an optimizer but does not specify software libraries, frameworks (e.g., PyTorch, TensorFlow), or their version numbers. |
| Experiment Setup | Yes | Density Ratio Estimator (EIM) / Variational function V (x) (f-GAN): 3 fully connected layers, 50 neurons each, trained with L2 regularization with factor 0.001, early stopping and batch size 1, 000 Updates EIM: MORE-like updates with ϵ = 0.05 for components and coefficients, 1, 000 samples per component and update Updates FGAN: Iterate single update steps for generator and discriminator using learning rates of 1e 3 and batch size of 1, 000. |