Autoinverse: Uncertainty Aware Inversion of Neural Networks

Authors: Navid Ansari, Hans-peter Seidel, Nima Vahidi Ferdowsi, Vahid Babaei

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

Reproducibility Variable Result LLM Response
Research Type Experimental We verify our proposed method through addressing a set of real-world problems in control, fabrication, and design. We evaluate the performance of Autoinverse through experimenting with the existing neural inverse methods and their uncertainty-aware counterparts.
Researcher Affiliation Academia Navid Ansari Max Planck Institute for Informatics Saarbrücken, Germany nansari@mpi-inf.mpg.de Hans-Peter Seidel Max Planck Institute for Informatics Saarbrücken, Germany hpseidel@mpi-sb.mpg.de Nima Vahidi Ferdowsi Max Planck Institute for Informatics Saarbrücken, Germany nvahidi@mpi-inf.mpg.de Vahid Babaei Max Planck Institute for Informatics Saarbrücken, Germany vbabaei@mpi-inf.mpg.de
Pseudocode No The paper describes methods using equations and prose, but it does not include formal pseudocode or algorithm blocks.
Open Source Code Yes Our code and data are available at: https://gitlab.mpi-klsb.mpg.de/nansari/autoinverse
Open Datasets Yes Our code and data are available at: https://gitlab.mpi-klsb.mpg.de/nansari/autoinverse. The training data consists of 10,000 pairs of samples generated by randomly sampling the NFP (Multi-joint robot). All networks in the ensemble NFP are trained on 40,000 printed patches [1] (Spectral printer). The training data consists of 50,000 samples queried by random sampling the actuation with an expansion ratios between -0.2 and 0.2 [33] (Soft robot).
Dataset Splits No The paper states 'Typically, we use 10% of the target performance for tuning our inverse methods,' implying a tuning/validation phase, but it does not provide explicit train/validation/test dataset splits (percentages or counts) for the underlying data samples themselves.
Hardware Specification No The paper does not provide specific hardware details such as CPU/GPU models, memory, or other computational specifications used for running experiments.
Software Dependencies No The paper does not explicitly list specific software dependencies with their version numbers required to reproduce the experiments.
Experiment Setup Yes We introduce α and β as hyperparameters to adjust the relative significance of aleatoric and epistemic uncertainties, respectively. We tune these parameters for 3 different sets of values for {α, β}: {{0.1, 1}, {1, 10}, {10, 100}}. All methods except MINI have around 3 million parameters. We inject Gaussian noise N(0, 0.1) to the spectrum of the samples with more than 0.4 LC density.