Are Disentangled Representations Helpful for Abstract Visual Reasoning?

Authors: Sjoerd van Steenkiste, Francesco Locatello, Jürgen Schmidhuber, Olivier Bachem

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this paper, we conduct a large-scale study that investigates whether disentangled representations are more suitable for abstract reasoning tasks. Using two new tasks similar to Raven s Progressive Matrices, we evaluate the usefulness of the representations learned by 360 state-of-the-art unsupervised disentanglement models. Based on these representations, we train 3600 abstract reasoning models and observe that disentangled representations do in fact lead to better down-stream performance.
Researcher Affiliation Collaboration Sjoerd van Steenkiste IDSIA, USI, SUPSI sjoerd@idsia.ch Francesco Locatello ETH Zurich, MPI-IS locatelf@ethz.ch Jürgen Schmidhuber IDSIA, USI, SUPSI, NNAISENSE juergen@idsia.ch Olivier Bachem Google Research, Brain Team bachem@google.com
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Code is made available as part of disentanglement_lib at https://git.io/JelEv.
Open Datasets Yes We create two new visual abstract reasoning tasks similar to Raven s Progressive Matrices [61] based on two disentanglement data sets: d Sprites [27], and 3dshapes [42].
Dataset Splits Yes We train each of these models using a batch size of 32 for 100K iterations where each mini-batch consists of newly generated random instances of the abstract reasoning tasks. Similarly, every 1000 iterations, we evaluate the accuracy on 100 mini-batches of fresh samples.
Hardware Specification Yes Reproducing these experiments requires approximately 2.73 GPU years (NVIDIA P100).
Software Dependencies No The paper mentions that code is available as part of 'disentanglement_lib' but does not specify particular software dependencies with version numbers (e.g., Python, PyTorch/TensorFlow, CUDA versions).
Experiment Setup Yes We train each of these models using a batch size of 32 for 100K iterations where each mini-batch consists of newly generated random instances of the abstract reasoning tasks. We first sample a set of 10 hyper-parameter configurations from our search space and then trained WRe N models using these configurations... The remaining experimental details are presented in Appendix A.