Concept Learners for Few-Shot Learning

Authors: Kaidi Cao, Maria Brbic, Jure Leskovec

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our model on few-shot tasks from diverse domains, including finegrained image classification, document categorization and cell type annotation on a novel dataset from a biological domain developed in our work. COMET significantly outperforms strong meta-learning baselines, achieving 6 15% relative improvement on the most challenging 1-shot learning tasks, while unlike existing methods providing interpretations behind the model s predictions.
Researcher Affiliation Academia Kaidi Cao , Maria Brbi c , Jure Leskovec Department of Computer Science Stanford University {kaidicao, mbrbic, jure}@cs.stanford.edu
Pseudocode No The paper describes the proposed method in Section 2, including mathematical formulations for concept learners and probability calculations. However, it does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code is publicly available at https://github.com/snap-stanford/comet.
Open Datasets Yes We develop a novel single-cell transcriptomic dataset based on the Tabula Muris dataset (Consortium, 2018; 2020) that comprises 105,960 cells of 124 cell types collected across 23 organs of the mouse model organism... This novel dataset along with the cross-organ evaluation splits is publicly available at https://snap.stanford.edu/comet. We use bird classification CUB-200-2011 (Wah et al., 2011) and flower classification Flowers-102 (Nilsback & Zisserman, 2008) datasets... In NLP domain, we apply COMET to benchmark document classification dataset Reuters (Lewis et al., 2004).
Dataset Splits Yes On the CUB dataset, we followed the evaluation protocol in (Chen et al., 2019b) and split the dataset into 100 base, 50 validation, and 50 test classes in the exactly same split. On the Tabula Muris, we use 15 organs for training, 4 organs for validation, and 4 organs for test, resulting into 59 base, 47 validation, and 37 test classes corresponding to cell types. The 102 classes of Flowers dataset are split into 52, 25, 25 as the training, validation and testing set, respectively. As for Reuters dataset, we leave out 5 classes for validation and 5 for test.
Hardware Specification No The paper mentions the use of a "four-layer convolutional backbones Conv-4" and a "simple backbone network structure containing two fully-connected layers". These describe model architectures, not the specific hardware (e.g., GPU model, CPU type) used to train or run the experiments. No specific hardware details are provided.
Software Dependencies No The paper mentions using "Adam optimizer (Kingma & Ba, 2014)". While it refers to a specific optimizer, it does not specify any software libraries (e.g., PyTorch, TensorFlow) or their version numbers, nor any other ancillary software dependencies with versions.
Experiment Setup Yes We train the 5-shot tasks for 40,000 episodes and 1-shot tasks for 60,000 episodes (Chen et al., 2019b). We use the Adam optimizer (Kingma & Ba, 2014) with an initial learning rate of 10 3 and weight decay 0.