Towards Context-Agnostic Learning Using Synthetic Data

Authors: Charles Jin, Martin Rinard

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically validate our methods by training deep neural networks for a variety of real-world image classification tasks using only a single synthetic example of each class, obtaining robust performance in the context-agnostic setting on natural data. Conversely, we find that classifiers trained without our techniques using only natural data achieve negligible accuracy even under relatively benign perturbations that leave a well-defined object in the foreground completely untouched.
Researcher Affiliation Academia Charles Jin MIT Cambridge, MA 02139 ccj@csail.mit.edu Martin Rinard MIT Cambridge, MA 02139 rinard@csail.mit.edu
Pseudocode Yes Algorithm 1: Greedy Bias Correction
Open Source Code No The paper uses PyTorch and mentions a PyTorch library for CAM methods (jacobgil/pytorch-grad-cam) in its references. However, it does not explicitly state that the source code for *their* proposed methodology is publicly available, nor does it provide a link to *their* repository.
Open Datasets Yes The target dataset is the German Traffic Sign Recognition Benchmark (GTSRB) [Stallkamp et al., 2012]... The target dataset, MNIST [Le Cun], consists of 60,000 training and 10,000 test images... We use the Omniglot [Lake et al., 2015] challenge...
Dataset Splits No The paper mentions training and testing datasets (e.g., '60,000 training and 10,000 test images' for MNIST). However, it does not explicitly specify the size, percentage, or method for a validation split in the main text.
Hardware Specification Yes Our implementation is written in Python (Paszke et al., 2019) using the PyTorch deep learning framework, and all experiments were carried out on a single NVIDIA 2080 Ti GPU.
Software Dependencies No Our implementation is written in Python (Paszke et al., 2019) using the PyTorch deep learning framework... While PyTorch is cited with a year, the specific version number for PyTorch, Python, or other libraries like OpenCV (mentioned in references) or scikit-image is not provided.
Experiment Setup Yes Appendix C provides the full experimental setup and training details. Our models are trained using the Adam optimizer with default parameters ( = 0.9, = 0.999), and a batch size of 10 for all datasets, for 100 epochs.