Countering Language Drift with Seeded Iterated Learning
Authors: Yuchen Lu, Soumye Singhal, Florian Strub, Aaron Courville, Olivier Pietquin
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate SIL in a toy-setting Lewis Game, and then scale it up to the translation game with natural language. In both settings, SIL helps counter language drift as well as it improves the task completion compared to baselines. |
| Researcher Affiliation | Collaboration | 1Mila, University of Montreal 2Deep Mind 3Google Research Brain Team 4CIFAR Fellow. |
| Pseudocode | Yes | Algorithm 1 Seeded Iterate Learning for S/R Games |
| Open Source Code | Yes | 1https://github.com/JACKHAHA363/langauge_ drift_lewis_game 2https://github.com/JACKHAHA363/ translation_game_drift |
| Open Datasets | Yes | First, they are independently pretrained on the IWSLT dataset (Cettolo et al., 2012) to learn the initial language distribution. The agents are then finetuned with interactive learning by sampling new translation scenarios from the Multi30k dataset (Elliott et al., 2016) |
| Dataset Splits | No | For Lewis game, we use 2500 training objects to train models. We have 3125 unique objects. We held out 625 objects for testing. The paper does not explicitly mention a separate validation set with specific split sizes or percentages for hyperparameter tuning, only train and test/held-out splits. |
| Hardware Specification | No | The paper mentions "computations support provided by Compute Canada" in the Acknowledgement section but does not provide specific details about the hardware (e.g., CPU, GPU models, memory) used for the experiments. |
| Software Dependencies | No | The paper mentions using "Moses tokenizer (Koehn et al., 2007)", "Sennrich subword-nmt toolkit (Sennrich et al., 2016)", and that the "codebase is based on Pytorch (Paszke et al., 2019)", but it does not specify version numbers for these software components or libraries. |
| Experiment Setup | Yes | For SIL: k1 = 1000, k2 = k 2 = 400. We used batch size 32 for the Lewis game. We used Adam (Kingma & Ba, 2014) with learning rate 0.0001. For the translation game, we set the number of recurrent units to 256 for both encoder and decoder. The dimension of the word embedding is 256, and we use a maximum sequence length of 50. |