Variational Inference of Disentangled Latent Concepts from Unlabeled Observations

Authors: Abhishek Kumar, Prasanna Sattigeri, Avinash Balakrishnan

ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically observe significant improvement over existing methods in terms of both disentanglement and data likelihood (reconstruction quality).
Researcher Affiliation Industry Abhishek Kumar, Prasanna Sattigeri, Avinash Balakrishnan IBM Research AI Yorktown Heights, NY {abhishk,psattig,avinash.bala}@us.ibm.com
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper does not provide any statement or link indicating that the source code for the methodology is openly available.
Open Datasets Yes We evaluate our proposed method, DIP-VAE, on three datasets (i) Celeb A (Liu et al., 2015): It consists of 202, 599 RGB face images of celebrities... (ii) 3D Chairs (Aubry et al., 2014): It consists of 1393 chair CAD models... (iii) 2D Shapes (Matthey et al., 2017): This is a synthetic dataset of binary 2D shapes...
Dataset Splits No The paper specifies training and testing splits (e.g., "90% for training and 10% for test" for Celeb A, and "First 80%... for training and the rest... for test" for 3D Chairs) but does not mention a separate validation split.
Hardware Specification No The paper does not provide any specific details about the hardware used for running the experiments (e.g., GPU/CPU models, memory specifications).
Software Dependencies No The paper mentions using CNN network architectures and a linear SVM, but it does not specify any software names with version numbers (e.g., specific deep learning frameworks like PyTorch or TensorFlow, or library versions like scikit-learn).
Experiment Setup Yes Hyperparameters. For the proposed DIP-VAE-I, in all our experiments we vary λod in the set {1, 2, 5, 10, 20, 50, 100, 500} while fixing λd = 10λod for 2D Shapes and 3D Chairs, and λd = 50λod for Celeb A. For DIP-VAE-II, we fix λod = λd for 2D Shapes, and λod = 2λd for Celeb A. Additionally, for DIP-VAE-II we also penalize the ℓ2-norm of third order central moments of qφ(z) with hyperparameter λ3 = 200 for 2D Shapes data (λ3 = 0 for Celeb A). For β-VAE, we experiment with β = {1, 2, 4, 8, 16, 25, 32, 64, 100, 128 , 200, 256} (where β = 1 corresponds to the VAE). We used a batch size of 400 for all 2D Shapes experiments and 100 for all Celeb A experiments.