Compositional Generalization from First Principles

Authors: Thaddäus Wiedemer, Prasanna Mayilvahanan, Matthias Bethge, Wieland Brendel

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

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our theory in a range of synthetic experiments and perform several ablation studies that relate our findings to empirical methods (Section 4).We validate our theoretical framework on the multi-sprite data. All models were trained for 2000 epochs on training sets of 100k samples using an NVIDIA RTX 2080 Ti; all test sets contain 10k samples. Table 1 summarizes the reconstruction quality achieved on the in-domain (ID) test set (P) and the entire latent space (Q) for all experiments.
Researcher Affiliation Academia 1University of Tübingen 2Tübingen AI Center 3Max-Planck-Institute for Intelligent Systems, Tübingen
Pseudocode No The paper includes schematics of models in Figure 6, but no explicit pseudocode or algorithm blocks are provided.
Open Source Code Yes Code available at https://github.com/brendel-group/compositional-ood-generalization
Open Datasets Yes We validate our theoretical framework on the multi-sprite data.We additionally conduct experiments on the CLEVR dataset [35], a popular benchmark for compositional generalization and object-centric learning.
Dataset Splits No The paper mentions training sets of 100k samples and test sets of 10k samples, and for CLEVR, setting aside 10% of ID samples for evaluation. However, it does not explicitly define a separate validation dataset split or a specific methodology for it.
Hardware Specification Yes All models were trained for 2000 epochs on training sets of 100k samples using an NVIDIA RTX 2080 Ti
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, or TensorFlow versions).
Experiment Setup Yes All models were trained for 2000 epochs on training sets of 100k samples...For training stability, the composition function is implemented as a soft pixel-wise addition using the sigmoid function σ( ) as x = σ( x1) x1 + σ( x1) x2.Both models are trained on samples (z, x) from the training set using an MSE reconstruction loss.