GlanceNets: Interpretable, Leak-proof Concept-based Models

Authors: Emanuele Marconato, Andrea Passerini, Stefano Teso

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we present results on several tasks showing that Glance Nets outperform CBNMs [20] in terms of alignment and robustness to leakage, while achieving comparable prediction accuracy. All experiments were implemented using Python 3 and Pytorch [51] and run on a server with 128 CPUs, 1Ti B RAM, and 8 A100 GPUs.
Researcher Affiliation Academia Emanuele Marconato Department of Computer Science University of Pisa & University of Trento Pisa, Italy emanuele.marconato@unitn.it Andrea Passerini Department of Computer Science University of Trento Trento, Italy andrea.passerini@unitn.it Stefano Teso Department of Computer Science University of Trento Trento, Italy stefano.teso@unitn.it
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks. The methods are described in prose and supported by architectural diagrams.
Open Source Code Yes The code is available at https://github.com/ema-marconato/glancenet.
Open Datasets Yes Data sets. We carried out our evaluation on two data sets taken from the disentanglement literature and a very challenging real-world data set. d Sprites [27] consists of 64 64 black-and-white images of sprites on a flat background... MPI3D [53] consists of 64 64 RGB rendered images of 3D shapes held by a robotic arm... Celeb A [28] is a collection of 178 218 RGB images of over 10k celebrities...
Dataset Splits Yes For d Sprites and MPI3D, we used a random 80/10/10 train/validation/test split, while for Celeb A we kept the original split [28].
Hardware Specification Yes All experiments were implemented using Python 3 and Pytorch [51] and run on a server with 128 CPUs, 1Ti B RAM, and 8 A100 GPUs.
Software Dependencies No All experiments were implemented using Python 3 and Pytorch [51]... Glance Nets were implemented on top of the disentanglement-pytorch [52] library. The paper mentions Python 3, but does not provide specific version numbers for Pytorch or the disentanglement-pytorch library, only citing their original papers from 2019.
Experiment Setup Yes For d Sprites and MPI3D, we implemented the encoder as a six layer convolutional neural net, while for Celeb A we adapted the convolutional architecture of Ghosh et al. [55]. We employed a six layer convolutional architecture for the decoder in all cases... For each data set, we chose the latent space dimension as the total number of generative factors... In particular, we used 7 latent factors for d Sprites, 21 for MPI3D and 10 for Celeb A. For d Sprites and MPI3D, we used a random 80/10/10 train/validation/test split.