Locality and Compositionality in Zero-Shot Learning
Authors: Tristan Sylvain, Linda Petrini, Devon Hjelm
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The results of our experiments show how locality, in terms of small parts of the input, and compositionality, i.e. how well can the learned representations be expressed as a function of a smaller vocabulary, are both deeply related to generalization and motivate the focus on more local-aware models in future research directions for representation learning. |
| Researcher Affiliation | Collaboration | Tristan Sylvain Mila, Universit e de Montr eal Montreal, Canada tristan.sylvain@gmail.com Linda Petrini University of Amsterdam Amsterdam, Netherlands lindapetrini@gmail.com R Devon Hjelm Microsoft Research, Mila Redmond, USA devon.hjelm@microsoft.com |
| Pseudocode | No | The paper describes methods and algorithms in prose and uses diagrams to illustrate concepts, but it does not include any formal pseudocode blocks or clearly labeled algorithm listings. |
| Open Source Code | No | All models used in this paper have been implemented in Py Torch, and code will be made publically available. |
| Open Datasets | Yes | We focus on three common ZSL dataset that allow us to explore compositionality and locality, namely Animals with Attributes 2 (Aw A2, Xian et al., 2018), Caltech-UCSD-Birds-200-2011 (CUB, Wah et al., 2011), SUN Attribute (SUN, Patterson & Hays, 2012). ... The details of each dataset are in the Appendix (Table 1). |
| Dataset Splits | Yes | Evaluation Protocol We used the ZSL splits constructed in Xian et al. (2017), as they are the most commonly used in the literature. ... The details of each dataset are in the Appendix (Table 1) which provides #Train classes and #Test classes. |
| Hardware Specification | No | The paper does not specify the hardware used for experiments, such as particular GPU models (e.g., NVIDIA A100), CPU models, or cloud computing instance types. It only mentions general implementation details without hardware specifics. |
| Software Dependencies | No | The paper states that "All models used in this paper have been implemented in Py Torch" but does not specify the version number of PyTorch or any other software libraries used, which is necessary for reproducibility. |
| Experiment Setup | Yes | All images were resized to size 128 128, random crops with aspect ratio 0.875 were used during training, and center crops with the same ratio were used during test. ... All models are optimized with Adam with a learning rate of 0.0001 with a batch size of 64. The final output of the encoder was chosen to be 1024 across all models. All the implementation details are available in the Appendix, in Section B. |