Restricting the Flow: Information Bottlenecks for Attribution
Authors: Karl Schulz, Leon Sixt, Federico Tombari, Tim Landgraf
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We compare our method against ten baselines using three different metrics on VGG-16 and Res Net-50, and find that our methods outperform all baselines in five out of six settings. We evaluate against ten different baselines. First, we calculated the Sensitivity-n metric proposed by Ancona et al. (2018). Secondly, we quantified how well the object of interest was localized using bounding boxes and extend the degradation task proposed by Ancona et al. (2017). |
| Researcher Affiliation | Academia | Karl Schulz1* , Leon Sixt2*, Federico Tombari1, Tim Landgraf2 * contributed equally work done at the Freie Universit at Berlin Technische Universit at M unchen1 Freie Universit at Berlin2 Corresponding authors: karl.schulz@tum.de, leon.sixt@fu-berlin.de |
| Pseudocode | No | The paper describes the Per-Sample Bottleneck and Readout Bottleneck with diagrams (Figure 2 and 3) and textual descriptions of the steps, but it does not provide formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | For reproducibility, we share our source code and provide an easy-to-use* implementation. *https://github.com/Biorobotics Lab/IBA |
| Open Datasets | Yes | The readout network is trained on the training set of the ILSVRC12 dataset (Russakovsky et al., 2015) for E = 20 epochs. |
| Dataset Splits | Yes | For each n, we generate 100 different index sets T and test each on 1000 randomly selected images from the validation set. In total, we run the bounding box evaluation on 11,849 images from the Image Net validation set. Both averages are taken over the validation set. |
| Hardware Specification | Yes | We are also grateful to Nvidia for providing us with a Titan Xp and to ZEDAT for granting us access to their HPC system. |
| Software Dependencies | No | As neural network architectures, we selected the Res Net-50 (He et al., 2016) and the VGG-16 Simonyan & Zisserman (2014), using pretrained weights from the torchvision package (Paszke et al., 2017). Optimizer Adam (Kingma & Ba (2014)). For LRP, we used the open source implementation by M. B ohle and F. Eitel (B ohle et al., 2019). |
| Experiment Setup | Yes | Table 2: Hyperparameters for Per-Sample Bottleneck. The optimization objective of the bottleneck is LCE + βLI as given in equation 6. For the Per-Sample Bottleneck, we investigate β = 1/k, 10/k, 100/k. The Readout network is trained with the best performing value β = 10/k. For the Per-Sample Bottleneck, we do 10 iterations using the Adam optimizer (Kingma & Ba, 2014) with learning rate 1 to fit the mask α. Table 3: Hyperparameters for the Readout Bottleneck. |