Wasserstein Variational Inference
Authors: Luca Ambrogioni, Umut Güçlü, Yağmur Güçlütürk, Max Hinne, Marcel A. J. van Gerven, Eric Maris
NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We will now demonstrate experimentally the effectiveness and robustness of Wasserstein variational inference. We focused our analysis on variational autoecoding problems on the MNIST dataset. |
| Researcher Affiliation | Academia | Luca Ambrogioni* Radboud University l.ambrogioni@donders.ru.nl Umut Güçlü* Radboud University u.guclu@donders.ru.nl Ya gmur Güçlütürk Radboud University y.gucluturk@donders.ru.nl Max Hinne University of Amsterdam m.hinne@uva.nl Eric Maris Radboud University e.maris@donders.ru.nl Marcel A. J. van Gerven Radboud University m.vangerven@donders.ru.nl |
| Pseudocode | Yes | Algorithm 1 Sinkhorn Iterations. C: Cost matrix, t: Number of iterations, ϵ: Regularization strength 1: procedure SINKHORN(C, t, ϵ) 2: K = exp( C/ϵ), n, m = shape(C) 3: r = ones(n, 1)/n, c = ones(m, 1)/m, u0 = r, τ = 0 4: while τ t do 5: a = KT uτ Juxtaposition denotes matrix product 6: b = c/a "/" denotes entrywise division 7: uτ+1 = m/(Kb), τ = τ + 1 v = c/(KT ut), Sϵ t = sum(ut (K C)v) "*" denotes entrywise product 8: return Sϵ t |
| Open Source Code | No | The paper does not provide an explicit statement about the release of source code for the methodology or a link to a code repository. |
| Open Datasets | Yes | We focused our analysis on variational autoecoding problems on the MNIST dataset. |
| Dataset Splits | Yes | We evaluated three performance metrics: 1) mean squared reconstruction error in the latent space, 2) pixelwise mean squared reconstruction error in the image space and 3) sample quality estimated as the smallest Euclidean distance with an image in the validation set. [...] For each method, we optimized this parameter on the validation set. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU, GPU models, or memory) used for running the experiments. It only describes the neural network architectures. |
| Software Dependencies | No | The paper mentions "differentiable programming frameworks" but does not specify any software libraries or dependencies with version numbers, which are necessary for reproducibility. |
| Experiment Setup | Yes | The generative models were parametrized by three-layered fully connected networks (100-300-500-1568) with Relu nonlinearities in the hidden layers. Similarly, the variational models were parametrized by three-layered Re Lu networks (784-500-300-100). The cost functional of our Wasserstein variational autoencoder (see Eq. 25) had the weights w1, w2, w3 and w4 different from zero. Conversely, in this experiment w5 was set to zero, meaning that we did not use a f-divergence component. We refer to this model as 1111. We trained 1111 using t = 20 Sinkhorn iterations. |