‘Less Than One’-Shot Learning: Learning N Classes From M < N Samples

Authors: Ilia Sucholutsky,Matthias Schonlau9739-9746

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

Reproducibility Variable Result LLM Response
Research Type Experimental We both theoretically and empirically demonstrate feasibility. We also perform a case study to confirm that soft labels can be used to represent training sets using fewer prototypes than there are classes, achieving large increases in sample-efficiency over regular (hard-label) prototypes.
Researcher Affiliation Academia Ilia Sucholutsky, Matthias Schonlau Department of Statistics and Actuarial Science University of Waterloo Waterloo, Ontario, Canada isucholu@uwaterloo.ca
Pseudocode No The paper defines 'Definition 4: The distance-weighted soft-label prototype k Nearest Neighbors (SLa Pk NN) classification rule' with formal notation, but not in a pseudocode block format.
Open Source Code Yes Our code, the appendix, and a live web demo can be found at https://github.com/ilia10000/SLkNN
Open Datasets Yes Sucholutsky and Schonlau (2019) showed that it is possible to design a set of five soft-labelled synthetic images, sometimes called prototypes , that train neural networks to over 90% accuracy on the ten-class MNIST task.
Dataset Splits No The paper mentions the MNIST dataset but does not provide specific details on training, validation, or test splits. It focuses on the theoretical aspects of class separation rather than empirical model training specifics.
Hardware Specification No No specific hardware specifications (e.g., GPU/CPU models, memory) used for experiments were mentioned in the paper.
Software Dependencies No No specific software dependencies with version numbers were mentioned in the paper.
Experiment Setup No The paper focuses on theoretical derivations and conceptual models (SLa Pk NN) rather than providing specific hyperparameter values or detailed training configurations for empirical experiments.