VICE: Variational Interpretable Concept Embeddings

Authors: Lukas Muttenthaler, Charles Y. Zheng, Patrick McClure, Robert A. Vandermeulen, Martin N Hebart, Francisco Pereira

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

Reproducibility Variable Result LLM Response
Research Type Experimental We compare VICE with SPo SE over three different datasets. One of these datasets contains concrete objects, another consists of adjectives, and the third is composed of food items. Experimentally, we find that VICE rivals or outperforms the performance of SPo SE in modeling human behavior.
Researcher Affiliation Academia Lukas Muttenthaler Machine Learning Group Technische Universität Berlin BIFOLD Berlin, Germany Charles Y. Zheng Machine Learning Team, FMRI Facility National Institute of Mental Health Bethesda, MD, USA Patrick Mc Clure Department of Computer Science Naval Postgraduate School Monterey, CA, USA Robert A. Vandermeulen Machine Learning Group Technische Universität Berlin BIFOLD Berlin, Germany Martin N. Hebart Vision and Computational Cognition Group MPI for Human Cognitive and Brain Sciences Leipzig, Germany Francisco Pereira Machine Learning Team, FMRI Facility National Institute of Mental Health Bethesda, MD, USA
Pseudocode Yes Algorithm 1 VICE optimization for individual triplets for a single training epoch
Open Source Code Yes A Py Torch implementation of VICE featuring Continuous Integration is publicly available at https: //github.com/Lukas Mut/VICE.
Open Datasets Yes Data We performed experiments for three triplet datasets: THINGS (used to develop SPo SE [46, 19]), ADJECTIVES4, and FOOD [7] 5. See https://osf.io/jum2f/ and https://things-initiative.org/ for data.
Dataset Splits Yes THINGS and ADJECTIVES are each comprised of a large training dataset which contains no repeats... For each of these training datasets, 10% of the triplets are assigned to a predefined validation set... FOOD contains data for every possible triplet combination for 36 objects and repeats for some triplets. We partitioned this dataset into train (45%), validation (5%), and test (50%) sets with disjoint triplets.
Hardware Specification No This study utilized the high-performance computational capabilities of the Biowulf Linux cluster at the National Institutes of Health, Bethesda, MD (http://biowulf.nih.gov) and the Raven and Cobra Linux clusters at the Max Planck Computing & Data Facility, Garching, Germany (https://www.mpcdf.mpg.de/services/supercomputing/). These are general cluster names and URLs, but no specific hardware components (e.g., GPU models, CPU types, memory) are detailed within the text.
Software Dependencies No We implemented both SPo SE and VICE in Py Torch [34] using Adam [22] with η = 0.001. While PyTorch is mentioned, a specific version number for PyTorch or any other software dependency is not provided.
Experiment Setup Yes Experimental setup We implemented both SPo SE and VICE in Py Torch [34] using Adam [22] with η = 0.001. To find the best VICE hyperparameter combination, we performed a grid search over σspike, σslab, πspike, optimizing Equation 3 and evaluating each model on the validation set. All experiments were performed over 20 different random initializations.