Learning to Draw: Emergent Communication through Sketching
Authors: Daniela Mihai, Jonathon Hare
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We next present a series of experiments where we explore if it is possible for the two agents to learn to successfully communicate, and what factors affect human interpretation of the drawings. We report numerical results averaged across 10 seeds for models evaluated on test sets isolated from training. Sample sketches from one seed are shown, but an overlay of 10 seeds can be found in Appendix J. |
| Researcher Affiliation | Academia | Daniela Mihai Electronics and Computer Science The University of Southampton Southampton, UK adm1g15@soton.ac.uk Jonathon Hare Electronics and Computer Science The University of Southampton Southampton, UK jsh2@soton.ac.uk |
| Pseudocode | No | The paper describes the model and procedures in text and through a diagram (Figure 1) but does not include explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | Full code for the model and all experiments can be found at https://github.com/Ddaniela13/Learning To Draw. |
| Open Datasets | Yes | We explore the game setups described in Section 3.1 and train our agents to play the games using 96 96 photographs from the STL-10 dataset [5]. ... We present results for the OO-game different setup played with 128 128 Caltech-101 images [10]. ... To further analyse what is being captured by the models we train the agents in the original game setting (using both normal and stylized backbone weights) with images from the Celeb A dataset [32]... |
| Dataset Splits | No | Batch size is K + 1, where K is the number of distractors, for all experiments. Unless otherwise stated, training was performed for 250 epochs. ... We report numerical results averaged across 10 seeds for models evaluated on test sets isolated from training. The paper mentions training and testing but does not explicitly specify a validation split or its size/percentage. |
| Hardware Specification | Yes | A mixture of Nvidia GTX1080s, RTX2080s, Quadro RTX8000s, and an RTX-Titan was used for training the models. |
| Software Dependencies | No | The paper mentions specific architectures and optimizers (e.g., VGG16, Adam) and pre-trained weights (torchvision Image Net weights) but does not provide specific version numbers for software dependencies like Python, PyTorch, or other libraries used in implementation. |
| Experiment Setup | Yes | Optimisation of the parameters of both agents is performed using the Adam optimiser with an initial learning rate of 1 10 4 for all experiments. For efficiency, we train the model with batches of games where the sender is given multiple images which are converted to sketches and passed to the receiver which reuses the same set of photographs for each sketch in the batch (with each sketch targeting a different receiver photograph). The order of the targets with respect to the input image s sketches is shuffled every batch. Batch size is K + 1, where K is the number of distractors, for all experiments. Unless otherwise stated, training was performed for 250 epochs. |