Evaluating the AppMax Method on Approximated Regression DNNs

For the purposes of the double-blind review process, the authors have been omitted. This information will be restored in the final version upon acceptance of the related manuscript. Datasets Used & Attributions To train our models and examine their behavior in certain data points, we used the following datasets: California Housing Dataset: Based on Pace and Barry (1997) and 1990 US census data. YearPredictionMSD: Thierry Bertin-Mahieux et al. (2011), The Million Song Dataset. UTKFace: Zhang, Song, et al. (2017) Table of Contents datasets.zip captures the exact train/dev/test data splits and the order of UTKFace images note that the current implementation of splitting relies on seeded randomness in PyTorch and NumPy models.zip contains the three trained regression models experiments.zip provides raw experiment results as well as charts and tables built using the data intervals.txt contains the results used to create Table I in Section III (The Error Bounds For Weight Rounding) the remaining results and outputs are organized into directories: there is one directory per dataset (e.g., california) there are directories containing the individual experiment results (for example, file 8bit/point_0006.pt contains the results corresponding to the 8-bit quantized model evaluated on a certain data point in the test set) metadata.pt captures LP bounds and scaling constants (allowing the visualizations to be generated without loading the original dataset) directories with names ending in _outputs (e.g., common_outputs, 8bit_outputs) contain tables and figures generated using the results of the experiments there is a JSON counterpart for every binary .pt file so that the contents can be examined easily these implementation-specific terms occur in the files: sample (meaning data point), nearby (polytope), union (extended polytope)