On Episodes, Prototypical Networks, and Few-Shot Learning
Authors: Steinar Laenen, Luca Bertinetto
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct a wide range of ablative experiments with Matching and Prototypical Networks, two of the most popular methods that use nonparametric approaches at the level of the episode. Their non-episodic counterparts are considerably simpler, have less hyperparameters, and improve their performance in multiple few-shot classification datasets. We conduct our experiments on mini Image Net [48], CIFAR-FS [5] and tiered Image Net [34], using the popular Res Net-12 variant first adopted by Lee et al. [24] as embedding function fθ 3. |
| Researcher Affiliation | Collaboration | Steinar Laenen School of Informatics University of Edinburgh V.S.E.Laenen@sms.ed.ac.uk Luca Bertinetto Five AI luca.bertinetto@five.ai |
| Pseudocode | No | The paper describes algorithms using mathematical formulations and prose, but no structured pseudocode or algorithm blocks are provided. |
| Open Source Code | Yes | Py Torch code is available at https://github.com/fiveai/on-episodes-fsl. |
| Open Datasets | Yes | We conduct our experiments on mini Image Net [48], CIFAR-FS [5] and tiered Image Net [34] |
| Dataset Splits | Yes | As standard [22], performance is assessed on episodes of 5-way, 15-query and 1or 5-shot. Each model is evaluated on 10,000 episodes sampled from the validation set during training, or from the test set during testing. |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, processor types, memory amounts, or detailed computer specifications) used for running the experiments are mentioned in the paper. |
| Software Dependencies | No | The paper mentions 'Py Torch code' but does not provide specific version numbers for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | A detailed description of benchmarks, architecture and choice of hyperparameters is deferred to Appendix F, while below we discuss the most important choices of the experimental setup. We define an episode by its number of shots n, the batch size b and the total number of images per class m + n (the sum of elements across support and query set). |