Labeling Neural Representations with Inverse Recognition

Authors: Kirill Bykov, Laura Kopf, Shinichi Nakajima, Marius Kloft, Marina Höhne

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the performance of the proposed approach across various datasets and models, and illustrate its practical use through multiple examples.
Researcher Affiliation Academia Kirill Bykov UMI Lab ATB Potsdam Potsdam, Germany kbykov@atb-potsdam.de Laura Kopf UMI Lab ATB Potsdam Potsdam, Germany lkopf@atb-potsdam.de Shinichi Nakajima Machine Learning Group TU Berlin Berlin, Germany nakajima@tu-berlin.de Marius Kloft Machine Learning Group RPTU Kaiserslautern-Landau Kaiserslautern, Germany kloft@cs.uni-kl.de Marina M.-C. Höhne UMI Lab ATB Potsdam University of Potsdam, Germany mhoehne@atb-potsdam.de
Pseudocode No Appendix A.3 'INVERT algorithm' describes the steps of the algorithm in paragraph form and numbered lists, but it does not present it as a formal pseudocode block or algorithm figure.
Open Source Code Yes 2The code can be accessed via the following link: https://github.com/lapalap/invert.
Open Datasets Yes For this, we employed the validation set of Image Net2012 [52] as the dataset DI in the INVERT process.
Dataset Splits Yes For this, we employed the validation set of Image Net2012 [52] as the dataset DI in the INVERT process.
Hardware Specification Yes Figure 5 showcases the running time of applying INVERT and Compositional Explanations for explaining 2048 neurons in layer 4 of the FCOS-Res Net50-FPN model [65] pre-trained on the MS COCO dataset [60] on a singe Tesla V100S-PCIE-32GB GPU.
Software Dependencies No The paper mentions using 'Py Torch models' (e.g., 'Res Net18 and Dense Net161 Py Torch models') but does not specify version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes Parameters of the proposed method include predetermined length L N, the beam size B N. Additionally, during the search process explanations could be constrained to the condition T(ϕ(C)) [α, β], where 0 α < β 0.5. In our experiments, unless otherwise specified, the parameter β is set to 0.5.