GAMR: A Guided Attention Model for (visual) Reasoning
Authors: Mohit Vaishnav, Thomas Serre
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on an array of visual reasoning tasks and datasets demonstrate GAMR s ability to learn visual routines in a robust and sample-efficient manner. |
| Researcher Affiliation | Academia | 1 Artificial and Natural Intelligence Toulouse Institute, Université de Toulouse, France 2 Carney Institute for Brain Science, Dpt. of Cognitive Linguistic & Psychological Sciences Brown University, Providence, RI 02912 3 Centre de Recherche Cerveau et Cognition, CNRS, Université de Toulouse, France |
| Pseudocode | Yes | Algorithm 1 Guided Attention Model for (visual) Reasoning (GAMR). |
| Open Source Code | No | The paper does not provide an explicit statement about releasing the source code for GAMR, nor does it provide a link to a repository containing their implementation code. It only links to the SVRT dataset generation code. |
| Open Datasets | Yes | This dataset can be generated with the code https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?p=svrt.git;a=summary provided by the SVRT authors with images of dimension 128 128. |
| Dataset Splits | Yes | We used unique sets of 4k and 40k samples for validation and test purposes. |
| Hardware Specification | No | The paper mentions 'Computing resources used supported by the Center for Computation and Visualization (NIH S10OD025181) at Brown and CALMIP supercomputing center (Grant 2016-p20019, 2016-p22041) at Federal University of Toulouse Midi-Pyrénées.' This describes general computing centers but does not provide specific hardware details such as GPU/CPU models or memory. |
| Software Dependencies | No | The paper mentions using 'Adam (Kingma & Ba, 2014) optimizer' and 'Optuna (Akiba et al., 2019) to get the best learning rates and weight decays'. While specific software/frameworks are named, their precise version numbers are not provided, nor are other key libraries or dependencies with versions. |
| Experiment Setup | Yes | We trained the model from scratch for a maximum of 100 epochs with an early stopping criterion of 99% accuracy on the validation set as followed in Vaishnav et al. (2022) using Adam (Kingma & Ba, 2014) optimizer and a binary cross-entropy loss. We used a hyperparameter optimization framework Optuna (Akiba et al., 2019) to get the best learning rates and weight decays for these tasks and reports the test accuracy for the models that gave the best validation scores. |