Weakly Supervised Representation Learning with Sparse Perturbations

Authors: Kartik Ahuja, Jason S. Hartford, Yoshua Bengio

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conducted two sets of experiments low-dimensional synthetic and image-based inputs that follow the DGP in equation (24). In the low-dimensional synthetic experiments we experimented with two choices for PZ a) uniform distribution with independent latents, b) normal distribution with latents that are blockwise independent (with block length d/2). We used an invertible multi-layer perceptron (MLP) (with 2 hidden layers) from Zimmermann et al. (2021) for g. We evaluated for latent dimensions d {6, 10, 20}. The training and test data size was 10000 and 5000 respectively. For the image-based experiments we used Py Game (Shinners, 2011) s rendering engine for g and generated 64 64 pixel images that look like those shown in Figure 1. The coordinates of each ball, zi, were drawn independently from a uniform distribution, zi U(0.1, 0.9). We varied the number of balls from 2 (d = 4) to 4 (d = 8).
Researcher Affiliation Academia Mila Quebec AI Institute, Université de Montréal. CIFAR Fellow. Correspondence to: kartik.ahuja@mila.quebec.
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Finally, the code to reproduce the experiments presented above can be found at https://github.com/ahujak/WSRL.
Open Datasets No For the image-based experiments we used Py Game (Shinners, 2011) s rendering engine for g and generated 64 64 pixel images that look like those shown in Figure 1. The coordinates of each ball, zi, were drawn independently from a uniform distribution, zi U(0.1, 0.9). ... In the low-dimensional synthetic experiments we experimented with two choices for PZ a) uniform distribution with independent latents, b) normal distribution with latents that are blockwise independent (with block length d/2)." The paper describes generating its own experimental data rather than using a pre-existing publicly available dataset for which concrete access information is provided.
Dataset Splits No The training and test data size was 10000 and 5000 respectively. For the image-based experiments... instead the images are generated online and we trained to convergence." The paper specifies training and test data sizes but does not explicitly mention a separate validation split or its size/percentage.
Hardware Specification Yes The experiments were conducted on a machine with 8 NVIDIA GeForce RTX 3090 GPUs. Each run took approximately 12 hours for the image-based experiments. For the low-dimensional synthetic experiments, each run took approximately 3 hours.
Software Dependencies No The code was implemented in PyTorch (Paszke et al., 2019) and Python 3.9." The paper mentions Python 3.9, which has a specific version, but for PyTorch, it only provides a citation without a specific version number.
Experiment Setup Yes For the low-dimensional synthetic experiments, we used Adam optimizer (Kingma and Ba, 2014) with a learning rate of 0.001. We trained for 100 epochs with a batch size of 256. For the image-based experiments, we used Adam optimizer with a learning rate of 0.0001. We trained to convergence with a batch size of 64.